AI Atlas

Autorska mapa tematów AI: obszary, tematy, krótkie opisy, przykłady i źródła.

A hand-curated map of AI topics. 1133 topics in 25 areas. Agents: start at llms.txt.

• AI w cyberbezpieczeństwie (AI for cybersecurity): Machine learning and language models in defending IT systems: intrusion and malware detection, vulnerability discovery, SOC work and AI-assisted attacks.
  • AI po stronie atakującego i testy penetracyjne (Offensive AI and AI-assisted penetration testing): Using models and agents for reconnaissance, exploitation and social engineering, both in legitimate penetration tests and in real attacks.
    • Agenci testów penetracyjnych (LLM penetration-testing agents): Language agents planning and executing penetration-test steps: reconnaissance, scanning, exploitation and reporting, supervised by a tester.
    • Ataki prowadzone z pomocą AI (AI-enabled and AI-orchestrated cyberattacks): Use of models by criminals and state groups: from writing convincing phishing to agents carrying out most stages of an intrusion.
  • AI w zespole reagowania (SOC) (AI for security operations (SOC)): Language models and agents supporting security analysts: alert triage, investigation, log analysis and threat-report analysis.
    • Analiza raportów o zagrożeniach (CTI) (Cyber threat intelligence analysis with language models): Language models extracting attack techniques, indicators of compromise and links to vulnerabilities from reports and bulletins in structured form.
      • Mapowanie opisów ataków na techniki MITRE ATT&CK (Automated mapping of threat text to ATT&CK techniques): Classifying sentences and threat reports into techniques and tactics from the MITRE ATT&CK catalogue, a multi-label text classification task.
    • Generowanie reguł detekcji z raportów o zagrożeniach (LLM-assisted detection rule generation): Using language models to turn attack descriptions and indicators from reports into detection rules in formats such as Sigma or YARA.
    • Triaż alertów i agenci śledczy (AI alert triage and investigation agents): Language agents that gather evidence from security tools for each alert, assess its credibility and prepare a justified verdict.
  • Analiza programów binarnych i inżynieria wsteczna z AI (AI for binary analysis and reverse engineering): Models trained on machine code and disassembly output that support decompilation, name recovery, similar-function search and malware analysis without source code.
    • Dekompilacja z użyciem modeli językowych (Neural and LLM-based decompilation): Translating machine code, or a classic decompiler's output, into readable, compilable source code with language models.
    • Odzyskiwanie nazw zmiennych w zdekompilowanym kodzie (Neural variable-name recovery in decompiled code): Models learn to restore meaningful variable names lost during compilation by using lexical and structural context from decompiled code.
    • Wyszukiwanie podobnych funkcji w kodzie binarnym (Binary code similarity detection): Learning embeddings of binary functions so the same code compiled with different compilers, options or architectures lies close together.
  • Szukanie i łatanie podatności z AI (AI-assisted vulnerability discovery and repair): Language models and agents that generate fuzz tests, read code for security bugs and propose fixes.
    • Bezpieczeństwo kodu pisanego przez AI (Security of AI-generated code): Measuring and reducing vulnerabilities in code produced by coding assistants and agents, via tests, benchmarks and secure-generation techniques.
    • Fuzzing wspomagany modelami językowymi (LLM-assisted fuzzing): Language models write and fix fuzzing harnesses and classify crashes, raising code coverage without manual work.
      • Generowanie sterowników fuzzingu przez modele językowe (LLM-based fuzz driver generation): Language models automatically writing the glue code that connects a fuzzer to a library's functions so random inputs reach deep into the tested code.
    • Systemy automatycznego znajdowania i łatania błędów (Cyber reasoning systems for automated find-and-patch): Autonomous systems combining fuzzing, static analysis and language models that find a vulnerability, confirm it and generate a tested patch.
      • Generowanie łatek na podatności przez modele językowe (LLM-based vulnerability patch generation): Generative models propose code changes intended to repair a known vulnerability. A candidate requires functional testing and security evaluation before use.
    • Wykrywanie podatności w kodzie przez modele językowe (Vulnerability detection in source code with language models): Language models and agents analysing source code and changes to flag security bugs and variants of known vulnerabilities.
  • Wykrywanie zagrożeń i włamań (Threat and intrusion detection): Models that detect attacks in network traffic, system logs and email: classifying known patterns and flagging departures from normal behaviour.
    • Analiza zachowań użytkowników i zagrożenia wewnętrzne (UEBA) (User and entity behaviour analytics (UEBA) and insider threat detection): Modelling the typical behaviour of accounts, users and devices in an organisation and detecting deviations that suggest a hijacked account or a malicious insider.
    • Wykrywanie anomalii w logach systemów (Log anomaly detection): Learning models of event sequences from system and application logs to detect unusual runs that signal failures or attacks.
      • Parsowanie logów do szablonów zdarzeń (Automated log parsing): Automatically splitting raw log messages into a fixed event template and variable parameters, as the first step in anomaly detection and diagnostics.
    • Wykrywanie ataków na grafach pochodzenia zdarzeń systemowych (Provenance-graph-based intrusion detection): Detecting long-running stealthy attacks by analysing dependency graphs between processes, files and network connections in an operating system.
    • Wykrywanie losowo generowanych domen (DGA) (Detection of algorithmically generated domains (DGA detection)): Classifying domain names as algorithmically generated by malware to reach its command server, often from the character string alone.
    • Wykrywanie phishingu (Phishing detection): Classifying messages, URLs and webpages as phishing from their content, URL representation and visual similarity to legitimate brands.
    • Wykrywanie włamań w ruchu sieciowym (Machine-learning network intrusion detection): Classifiers and anomaly detectors trained on network flows and packets, separating normal traffic from attacks.
  • Wykrywanie złośliwego oprogramowania (Machine-learning malware detection): Classifying files and program behaviour as malicious or benign from static and dynamic features.
    • Omijanie detektorów malware przykładami adwersarialnymi (Adversarial evasion of malware detectors): Modifying real malicious programs so they keep working while a classifier judges them benign, and defending against such attacks.
    • Starzenie się detektorów i realistyczna ocena w czasie (Concept drift and time-aware evaluation of malware detectors): Methods for evaluating and maintaining malware detectors that respect data chronology, realistic class ratios and performance decay over time.
    • Statyczna klasyfikacja plików wykonywalnych (Static malware classification of executables): Classifying executables (e.g. Windows PE files) as malicious or benign from features extracted without running the program.
    • Wykrywanie malware po zachowaniu w piaskownicy (Dynamic (behavioural) malware detection): Classifying programs from traces of their execution in an isolated environment, most often sequences of system API calls with their arguments.
• AI w edukacji (AI in education): AI applied to teaching and learning: learner modelling and learning analytics, tutoring systems and personalisation, and automated assessment and feedback.
  • Automatyczne ocenianie i informacja zwrotna (Automated assessment and feedback): Automated scoring of essays and short answers, feedback generation, and creation of test questions and distractors.
    • Automatyczna ocena wymowy (Automatic pronunciation assessment (CAPT)): Assessing language learners' pronunciation at the phone level and the prosodic level (stress, rhythm, intonation); the core of computer-assisted pronunciation training.
    • Automatyczna ocena wypracowań (Automated essay scoring): Predicting a holistic score or trait scores (coherence, argumentation, prompt adherence) for student-written texts.
      • Automatyczne komentarze do wypracowań (Automated writing evaluation and feedback (AWE)): Systems that generate feedback on learner writing (grammar, style, structure, and content), traditionally using rule-based and statistical methods; generative language models are being studied as a newer possibility.
    • Automatyczne sprawdzanie krótkich odpowiedzi (Automatic short-answer grading): Scoring short text answers against a reference answer across subjects and grading scales.
    • Automatyczne sprawdzanie zadań programistycznych (Automated grading and feedback for programming assignments): Tools that assess the correctness and selected quality characteristics of student code through dynamic analysis (such as unit tests), static analysis, or comparison with reference solutions.
    • Automatyczne tworzenie pytań i testów (Automatic question generation for education): Generating questions (open, cloze, multiple-choice) and distractors from instructional text, with control over difficulty and cognitive level.
    • Komputerowe testowanie adaptacyjne (CAT) (Computerized adaptive testing (CAT)): A testing procedure that draws each next item from a calibrated bank to give the most information at the current ability estimate.
  • Modelowanie ucznia i analityka uczenia się (Learner modelling and learning analytics): Models that estimate a learner's knowledge, engagement and risk of failure from answers, e-learning platform logs and study records.
    • Teoria odpowiedzi na pozycje testowe (IRT) (Item response theory (IRT)): A family of psychometric models describing the probability of a correct answer as a function of a person's latent ability and item parameters.
    • Wczesne ostrzeganie o uczniach zagrożonych niepowodzeniem (Early warning: predicting at-risk students): Classification and regression on LMS and student-record data to detect early the learners at risk of failing or dropping a course.
    • Wykrywanie zaangażowania i emocji ucznia (Learner engagement and affect detection): Detecting learner engagement and attempts to infer emotions from visual signals or behavior in a learning system.
    • Śledzenie wiedzy ucznia (knowledge tracing) (Knowledge tracing): Modelling skill mastery as it changes over time from sequences of correct and incorrect answers; from BKT to deep models.
      • Bayesowskie śledzenie wiedzy (BKT) (Bayesian knowledge tracing (BKT)): A hidden Markov model with a binary skill-mastery state and four parameters: prior knowledge, learning, guess and slip.
      • Głębokie śledzenie wiedzy (DKT) (Deep knowledge tracing (DKT)): A recurrent network that predicts the probability of answering upcoming questions correctly from the sequence of a learner's past interactions.
      • Śledzenie wiedzy z mechanizmem uwagi (SAKT, AKT) (Attention-based knowledge tracing (SAKT, AKT)): A family of knowledge-tracing models in which self-attention weights past interactions by their relevance to the current question.
  • Systemy tutorskie i nauka dopasowana do ucznia (Intelligent tutoring and adaptive learning): Systems that guide learners individually: classic ITS with domain and learner models, LLM-based tutors, support for live human tutors and review scheduling.
    • Automatyczne generowanie podpowiedzi (Automatic hint generation): Generating successive, increasingly specific hints in tutoring systems: from methods built on other learners' solutions to language models.
    • Korepetytorzy na modelach językowych z zabezpieczeniami pedagogicznymi (LLM tutors with pedagogical guardrails): LLM-based tutors steered by pedagogical instructions and constraints (hints instead of answers, prepared solutions, scaffolding), judged by their effect on learning.
    • Planowanie powtórek (spaced repetition) (Spaced repetition scheduling): Forgetting models trained on review history that predict, for each card, the probability of recall and set the next review date.
      • Modele zapominania uczone z danych (Learned memory models for spaced repetition (HLR, FSRS)): Models that estimate recall probability as a function of time and memory state, with parameters fitted to review logs; the basis of modern review schedulers.
    • Podpowiedzi AI dla nauczycieli i korepetytorów na żywo (Real-time AI guidance for human tutors): Human–AI systems in which a model, built on expert patterns, suggests strategies and replies to a tutor in real time.
• AI w finansach (AI in finance): Machine learning in banking, payments, investing and insurance: fraud detection, AML/KYC, credit scoring, quantitative trading, financial NLP and model risk.
  • AI w ubezpieczeniach (AI in insurance): Machine learning in insurance: pricing and underwriting, claims handling, fraud detection and customer service, under sector supervision and the AI Act.
    • Ocena szkód ze zdjęć w likwidacji szkód (Photo-based damage assessment for insurance claims): Detection and segmentation of damage types (dent, scratch, crack, shattered glass) in vehicle photos for claims processing.
    • Taryfikacja ubezpieczeń z uczeniem maszynowym (Machine-learning insurance pricing (actuarial learning)): Claim frequency and severity models for premium pricing: from generalised linear models to neural networks and tree ensembles.
  • Inwestowanie ilościowe i handel algorytmiczny (Quantitative investing and algorithmic trading): Machine learning in asset management: return prediction, portfolio construction, order execution and rigorous backtesting.
    • Optymalne wykonanie dużych zleceń (Optimal trade execution): Scheduling and pricing child orders to minimise the cost of executing a large order within a set time.
    • Przeuczenie backtestu i walidacja strategii (Backtest overfitting and strategy validation): Methods for detecting overfitting when many strategies are tested, including the probability of backtest overfitting (PBO) estimated with CSCV.
    • Przewidywanie stóp zwrotu modelami ML (Machine-learning return prediction): Predicting the cross-section of expected returns from hundreds of company characteristics with regularised regression, trees and neural networks.
    • Uczenie ze wzmocnieniem w handlu i alokacji portfela (Reinforcement learning for trading and portfolio allocation): Trading and portfolio allocation as an MDP, training deep RL agents (PPO, SAC, DDPG) in market environments with transaction costs.
    • Zespoły agentów językowych podejmujące decyzje inwestycyjne (LLM multi-agent trading systems): Systems where specialised LLM agents (analysts, debate, trader, risk) jointly make trading decisions from market and text data.
  • Ocena zdolności kredytowej (scoring) (Credit scoring and underwriting): Probability-of-default (PD) models and credit decisions: from scorecards to gradient boosting, under explainability and non-discrimination requirements.
    • Karty scoringowe i modele PD (Credit scorecards and PD models): Logistic regression on binned features encoded by weight of evidence (WoE), rescaled to points; the reference pattern for credit scoring.
    • Modele straty przy niewypłacalności (LGD) i oczekiwanej straty kredytowej (Loss given default (LGD) and expected credit loss models): LGD and EAD models that combine with PD into expected loss (EL = PD×LGD×EAD) for regulatory capital and IFRS 9 / CECL reserves.
    • Ocena kredytowa z historii rachunku (dane alternatywne) (Cash-flow underwriting with alternative data): Credit scoring on features from account transaction data (inflows, balances, obligations) obtained with consent, e.g. via open banking.
    • Wnioskowanie o odrzuconych wnioskach (reject inference) (Reject inference in credit scoring): Methods correcting sample selection bias in scoring by assigning inferred outcomes to rejected applications before model training.
    • Wyjaśnianie decyzji kredytowych i równe traktowanie (Explaining credit decisions and fair lending): Adverse-action reason codes, feature attribution (SHAP), counterfactual explanations and discrimination testing for regulated credit models.
  • Przeciwdziałanie praniu pieniędzy i weryfikacja klientów (AML/KYC) (Anti-money laundering and customer verification (AML/KYC)): Machine learning in financial-crime compliance: transaction monitoring for money laundering, flow-graph analysis and remote identity verification.
    • Analiza grafu transakcji (sieci kont i przepływów) (Transaction graph analysis): Modelling transactions as an account–flow graph and classifying nodes or edges with graph features and GNNs.
    • Monitorowanie transakcji i priorytetyzacja alertów AML (AML transaction monitoring and alert triage): Detecting suspicious patterns (structuring, fan-in, cycles) in transaction streams and ranking alerts for manual review.
    • Zdalna weryfikacja tożsamości i wykrywanie prób podszycia (eKYC) (Remote identity verification and liveness detection (eKYC)): Document verification, face-to-document matching and detection of presentation attacks and deepfakes during remote onboarding.
  • Przetwarzanie języka w finansach (Natural language processing in finance): NLP and LLMs on financial text: news sentiment analysis, extraction from reports, question answering that requires calculations over tables and text.
    • Analiza sentymentu wiadomości finansowych (Financial news sentiment analysis): Classifying the tone of headlines, announcements and earnings calls with models tuned to the financial domain.
    • Analiza telekonferencji wynikowych (tekst i głos) (Earnings call analysis from text and voice): Multimodal models combining transcripts and vocal features from earnings calls to predict company volatility and risk.
    • Automatyczne tagowanie liczb w sprawozdaniach (XBRL) (XBRL tagging of financial filings): Numeric entity recognition in filings, assigning XBRL taxonomy tags with hundreds of context-dependent types.
    • Modele językowe uczone na tekstach finansowych (Finance-specific language models): Language models pretrained or further trained on financial corpora (filings, news, transcripts) to improve domain task performance.
    • Pytania i obliczenia na raportach finansowych (Question answering and numerical reasoning over financial reports): Answering questions that require retrieval and multi-step calculation over tables and text in financial reports; FinQA, TAT-QA, FinanceBench benchmarks.
  • Wykrywanie oszustw płatniczych (Payment fraud detection): Real-time transaction classification on extremely imbalanced data, combining rules, learned models and analyst review.
    • Biometria behawioralna (sposób pisania i obsługi telefonu) (Behavioural biometrics for fraud prevention): Continuous authentication and session risk scoring from typing dynamics, mouse movement, touch and device sensors.
    • Ocena i walidacja modeli antyfraudowych (Evaluating and validating fraud models): Metrics and validation schemes for fraud detection: precision among top-k alerts, PR curves, time-based validation with label delay.
    • Ocena ryzyka transakcji w czasie rzeczywistym (Real-time transaction risk scoring): A scoring model on time-window aggregated features (customer, terminal, device) that returns a transaction's risk within milliseconds.
    • Uczenie na opóźnionych etykietach i zmieniających się oszustwach (Learning with delayed labels and concept drift in fraud detection): Streaming training of a fraud classifier from two label sources, immediate alert feedback and delayed labels, under concept drift.
    • Wykrywanie oszustw, w których ofiara sama zleca przelew (Authorised push payment (APP) scam detection): Models scoring the risk of customer-authorised transfers for social-engineering manipulation, at both the sending and receiving bank.
  • Zarządzanie ryzykiem modeli w instytucjach finansowych (Model risk management in financial institutions): The model lifecycle in financial institutions: inventory, materiality assessment, independent validation, monitoring and oversight; in 2026 US regulators replaced SR 11-7 with new guidance.
    • Nadzór nad modelami generatywnymi w bankach (Generative AI risk management in financial institutions): Extending model risk management to generative AI: AI inventory, materiality assessment, use-case testing, guardrails, human oversight and vendor controls.
    • Niezależna walidacja i monitoring modeli finansowych (Independent validation and ongoing monitoring of financial models): Independent assessment of a model's concept, data, implementation and outcomes plus ongoing performance monitoring, proportionate to model materiality.
• AI w grach (AI in games): AI applied to games: game-playing agents, procedural content generation, non-player character behaviour and player modelling; games also serve as AI research testbeds.
  • Agenci grający w gry (Game-playing agents): Agents reaching high skill in board, card and video games through search, reinforcement learning, self-play and visual understanding.
    • Agenci ogólni grający w wiele gier (Generalist game agents): Agents that take screen images and language instructions and act via keyboard and mouse across many 3D games, often built around a vision-language model.
    • Automatyczne testowanie gier przez agentów (Automated playtesting with agents): Using agents trained by reinforcement, imitation or exploration to cover the game space, find bugs and exploits, and assess difficulty.
    • Gry z ukrytą informacją (poker, CFR) (Imperfect-information games and counterfactual regret minimisation): Algorithms for hidden-information games built on iterative counterfactual regret minimisation, game abstraction and real-time search.
    • Uczenie grania z filmów z rozgrywki (VPT) (Video pretraining from gameplay videos): Semi-supervised imitation: an inverse dynamics model trained on a small labelled set annotates a large video corpus, then a policy is behaviour-cloned on it.
    • Uczenie przez grę z samym sobą (self-play) (Self-play and league training): Training agents through games against current or past versions of themselves, often with a population of strategies (a league) to prevent cyclic forgetting.
  • Animacja postaci sterowana danymi (Data-driven character animation): Techniques generating body and facial motion of characters from motion-capture data: frame search, neural motion controllers, generative motion models and audio-driven animation.
    • Animacja twarzy z nagrania głosu (Audio-driven facial animation): Models mapping a speech signal to a sequence of facial-animation parameters for a 3D character, followed by retargeting the motion to its rig.
    • Dopasowywanie ruchu z bazy nagrań (motion matching) (Motion matching): A kinematic animation method that every frame or few frames searches the recording database for the nearest pose-and-trajectory feature vector and jumps playback there.
    • Generowanie ruchu postaci z opisu tekstowego (Text-to-motion generation): Generative models, most often diffusion transformers, producing skeleton pose sequences conditioned on text, an action label or other signals.
    • Neuronowe sterowanie ruchem postaci (Neural character controllers): Autoregressive networks predicting the next frame's pose from the previous state, control input and scene geometry, often with weights that depend on the motion phase.
  • Modelowanie gracza (Player modelling): Building computational models of players' skill, preferences, emotions and behaviour from gameplay data, physiological signals and context.
    • Bayesowskie oceny umiejętności graczy (Bayesian player skill rating): Bayesian inference of latent skill and its uncertainty from individual or team outcomes; ratings can support balanced matchmaking.
    • Behawioralne wykrywanie oszustw w grach (Behavioural cheat detection in games): Classifying cheating programs from multivariate time series of player input, without access to in-game data.
    • Dynamiczne dopasowanie trudności (Dynamic difficulty adjustment (DDA)): Real-time tuning of game parameters to the player's skill or emotional state, estimated from gameplay performance or physiological signals.
    • Grupowanie graczy według stylu gry (Player segmentation by behavioural clustering): Unsupervised clustering of behavioural telemetry (k-means, archetypal analysis, matrix factorisation) into interpretable player profiles for game design and analytics.
    • Przewidywanie odejść graczy (Player churn prediction): Classification predicting whether a player will stop playing within a defined window, based on session telemetry and aggregated activity features.
  • Proceduralne generowanie treści gier (Procedural content generation (PCG)): Algorithmic creation of game content — levels, maps, quests, items, rules — using constructive methods, search, constraint solving or machine learning.
    • Generowanie poziomów uczone na przykładach (Procedural content generation via machine learning (PCGML)): Generating game content with models trained on existing content, including Markov chains, recurrent networks, autoencoders and generative networks.
    • Generowanie terenu z szumu i fraktali (Noise- and fractal-based terrain generation): Classic terrain PCG: heightmaps from gradient or value noise summed over octaves and subdivision algorithms (diamond-square) approximating fractional Brownian motion.
    • Generowanie treści przez przeszukiwanie (ewolucyjne) (Search-based procedural content generation): Generating game content with metaheuristics, mostly evolutionary, which search a space of content representations guided by a quality evaluation function.
    • Generowanie z gramatyk i L-systemów (Grammars and L-systems for content generation): Content generation by repeatedly applying production rules: parallel string rewriting (L-systems) for plants and graph grammars for missions and level layouts.
    • Generowanie z lokalnych wzorów (Wave Function Collapse) (Wave Function Collapse and constraint-based generation): Generating maps and textures by solving neighbourhood constraints learned from an example or given as a tile list, collapsing the lowest-entropy cells first.
    • Grywalne światy z modeli generatywnych (neuronowe silniki gier) (Neural game engines and interactive world generation): Generative models, usually diffusion or autoregressive, that predict the next frames of an interactive world from image history and player actions in real time.
    • Interaktywna fabuła i reżyser historii (Interactive narrative and drama management): Systems in which an experience manager tracks player actions and steers events and characters to reconcile player agency with plot coherence.
  • Zachowanie postaci w grach (NPC behaviour and game character AI): Techniques that drive non-player characters: state machines, behaviour trees, goal planning, pathfinding and characters driven by language models.
    • Automaty stanów dla postaci (FSM) (Finite state machines for game characters): A control model where a character is always in exactly one state and events trigger defined transitions; hierarchical and stack variants add sharing and memory.
    • Drzewa zachowań (Behaviour trees): A hierarchical, modular control structure in which sequence, selector and action nodes are ticked periodically and return success, failure or running.
    • Hierarchiczne planowanie zadań postaci (HTN) (Hierarchical task network planning for game characters): A planner that decomposes compound tasks into primitive ones via conditioned methods, simulating action effects on a copy of the world state.
    • Mapy wpływów do decyzji taktycznych (Influence maps for tactical decisions): A grid representation of space in which agents spread proximity and threat values that decay with distance; layers combined arithmetically serve tactical queries.
    • Planowanie działań postaci (GOAP) (Goal-oriented action planning for game characters): Real-time planning for characters in which the system searches for a sequence of actions described by preconditions, effects and costs.
    • Postacie rozmawiające dzięki modelom językowym (LLM-driven NPCs and generative agents): Characters and social simulations driven by language models with memory, reflection and planning, grounded in game rules and state.
    • Siatki nawigacyjne i odnajdywanie drogi (Navigation meshes and pathfinding): Representing walkable space as a mesh of convex polygons generated from level geometry, with pathfinding and collision avoidance on top.
    • Wybór działań przez punktację (utility AI) (Utility AI for game characters): A decision architecture in which every action receives a normalised score from weighted considerations passed through response curves, and the agent picks the highest-utility action.
• AI w medycynie i ochronie zdrowia (AI in healthcare and medicine): Machine learning for diagnosis, clinical documentation, risk prediction and patient monitoring, together with clinical validation and medical-device regulation.
  • Analiza obrazów medycznych (Medical image analysis): Classification, detection and segmentation on X-ray, CT, MRI, ultrasound, mammography and histopathology images, usually as decision support for clinicians.
    • Automatyczne opisy badań obrazowych (Radiology report generation): Generating the radiology report text (findings and impression) from a medical image with models that couple an image encoder to a language model.
    • Ocena znamion i zmian skórnych ze zdjęć (AI classification of skin lesions): Neural-network classification of clinical and dermoscopic skin-lesion images (e.g. melanoma versus naevus), with a focus on performance across skin tones.
    • Patologia cyfrowa: analiza skanów preparatów tkankowych (Computational pathology on whole-slide images): Analysis of digital whole-slide images (WSI): cancer classification, metastasis detection and biomarker prediction from tissue images.
      • Modele bazowe w patologii (Pathology foundation models): Tile and slide encoders pretrained self-supervised on millions of histopathology images, used as general feature extractors for pathology tasks.
      • Uczenie z etykiety całego preparatu (wiele instancji) (Multiple instance learning for whole-slide images): Weakly supervised learning in which a slide is a bag of tiles with one label, and attention pooling combines tile features into a slide-level prediction.
    • Przesiewowa ocena zdjęć dna oka (AI screening of retinal images (diabetic retinopathy)): CNN classification of fundus photographs to detect referable diabetic retinopathy, including autonomous operation without a human reader.
    • Segmentacja narządów i zmian na obrazach medycznych (Medical image segmentation): Assigning each voxel of a 2D or 3D image an organ or lesion label, for radiotherapy planning, volumetry and surgery among others.
    • Wykrywanie zmian na zdjęciach radiologicznych (Computer-aided detection and triage in radiology): CADe/CADx and triage systems that detect, localise and prioritise findings in radiology images, for example in screening mammography.
      • AI w przesiewowej mammografii (AI in mammography screening): Supporting screening mammography with models that score cancer risk, used to triage between single and double reading and to mark suspicious regions.
  • Język kliniczny i dokumentacja medyczna (Clinical language processing and documentation): NLP on clinical text: note generation, information extraction, diagnosis coding, de-identification and summarisation of patient records.
    • Automatyczne notatki z wizyty (ambient scribe) (Ambient clinical documentation (AI scribes)): Systems that turn a visit recording into a structured clinical note (e.g. SOAP) for the clinician to review and sign.
    • Kodowanie rozpoznań i wyciąganie danych z dokumentacji (Automated clinical coding and information extraction): Multi-label classification of clinical notes into ICD codes, plus extraction of clinical entities (diseases, drugs, doses) and mapping to terminologies.
    • Szkice odpowiedzi na wiadomości od pacjentów (AI-drafted replies to patient messages): Language-model generation of draft replies to patient messages in a portal integrated with the health record, for clinician editing and approval.
    • Usuwanie danych osobowych z dokumentacji medycznej (Clinical text de-identification): Detecting and removing or replacing protected health information (PHI) in clinical text, framed as named-entity recognition.
  • Koncepcja wszechstronnych modeli medycznych (Generalist medical AI (multimodal medical models)): A proposed foundation-model paradigm intended to combine many medical modalities (images, EHRs, results, genomics and text) and perform tasks specified by instructions.
  • Modele językowe w medycynie (Medical large language models): General and medically tuned LLMs used for question answering, work with clinical cases and health communication, together with evaluation of their accuracy and safety.
    • Ocena modeli na pytaniach i rozmowach medycznych (Evaluating LLMs on medical questions and conversations): Benchmarks and protocols for evaluating medical LLMs: from exam questions to multi-turn conversations scored with physician rubrics and harm review.
    • Odpowiedzi oparte na literaturze i zaleceniach medycznych (RAG) (Retrieval-augmented generation for medicine): RAG over medical corpora (literature, textbooks and guidelines) that supplies retrieved passages and can improve accuracy while making the basis of an answer easier to inspect.
    • Wspomaganie diagnozy różnicowej (LLM-assisted differential diagnosis): Using LLMs to generate a ranked differential-diagnosis list from a case description, standalone or as clinician support.
  • Prognozowanie ryzyka z dokumentacji pacjenta (Clinical risk prediction from health records): Diagnostic and prognostic models on electronic health record (EHR) data: early warning, complication risk, survival and readmissions.
    • Modele bazowe na danych z dokumentacji pacjenta (Foundation models for electronic health records): Models pretrained on sequences of structured EHR events (codes, medications, lab results) that produce patient representations for many prediction tasks.
    • Szacowanie, komu leczenie pomoże (efekt indywidualny) (Individualised treatment effect estimation): Estimating the conditional average treatment effect (CATE) from randomised or observational data with causal machine learning, as a basis for treatment recommendations.
    • Uczenie maszynowe dla czasu do zdarzenia (Machine learning for time-to-event outcomes): Models predicting time to an event with censored data, from the Cox model to survival forests and neural networks.
    • Wczesne ostrzeganie o pogorszeniu stanu (np. sepsa) (Early warning systems for clinical deterioration (e.g. sepsis)): Models running continuously on EHR data that flag patients at rising risk of sepsis or deterioration and alert staff.
  • Sprawdzanie objawów i kierowanie po pomoc (Symptom checkers and automated triage): Patient-facing tools that use answers from an interactive symptom interview to suggest possible diagnoses and the urgency of care.
  • Sygnały z ciała i urządzenia noszone (Physiological signals and wearables): Machine learning on physiological time series (ECG, PPG, EEG, polysomnography, wearable data) for detecting and monitoring disease.
    • Analiza EEG i wykrywanie napadów padaczkowych (Deep learning for EEG and seizure detection): Detecting seizures in intracranial EEG and epileptiform discharges in scalp EEG with convolutional, recurrent and transformer networks.
    • Analiza zapisu EKG przez sieci neuronowe (Deep learning for ECG analysis): Deep networks for classifying ECG recordings, especially rhythms and arrhythmias; other uses require separate validation.
    • Automatyczna ocena faz snu (Automated sleep staging): Classifying 30-second polysomnography epochs (EEG, EOG, EMG) into sleep stages under AASM rules with convolutional and recurrent networks.
    • Monitorowanie zdrowia z zegarków i opasek (Wearable health monitoring): Algorithms analysing consumer wearable data (PPG, accelerometer and single-lead ECG) for screening, trend tracking and selected remote-monitoring uses.
  • Walidacja kliniczna i dopuszczenie do użytku (Clinical validation and regulatory approval): Evaluating medical models: external validation and prospective studies, reporting standards, medical-device approval and post-deployment monitoring.
    • Dopuszczenie AI jako wyrobu medycznego (Regulation of AI as a medical device (SaMD)): Regulatory frameworks for AI software with a medical purpose: risk classes, approval pathways, change control plans and post-market surveillance.
    • Nadzór nad modelem po wdrożeniu w szpitalu (Post-deployment monitoring of clinical AI): Continuous checking of a clinical model's inputs, outputs and performance after deployment, with drift detection, subgroup analysis and a response procedure.
    • Przenoszenie modeli między szpitalami i ukryte skróty (Cross-site generalisation and shortcut learning in medical AI): Loss of medical model performance when the site, equipment or population changes, due to distribution shift and shortcut learning (confounders).
    • Sprawdzanie stronniczości modeli wobec grup pacjentów (Bias and equity audits of clinical algorithms): Assessing differences in clinical models' performance and consequences across patient groups and finding bias sources, including proxy labels.
    • Standardy raportowania badań nad AI w medycynie (Reporting guidelines for clinical AI studies): Reporting guidelines (TRIPOD+AI, CONSORT-AI, SPIRIT-AI, DECIDE-AI, TRIPOD-LLM) for studies of prediction models and AI interventions in medicine.
• AI w nauce (AI for science): Deep learning applied to scientific problems at quantum, atomistic and continuum scales: biomolecules, chemistry, materials, climate and physical equations.
  • Automatyzacja badań naukowych (AI-driven scientific discovery and research automation): Agentic LLM systems and optimisation loops automating the scientific cycle: hypothesis, plan, experiment (computational or robotic), analysis, report.
    • Agenci generujący hipotezy i eksperymenty obliczeniowe (Hypothesis-generating research agents): Multi-agent LLM systems that generate, evaluate and evolve hypotheses and, in computational fields, also write code, run experiments and draft papers.
      • Ideacja naukowa przez analogie (Analogy-based scientific ideation): An agent system decomposes ideation into gap finding, analogy-based innovation and proposal writing, trained with reinforcement learning.
    • Laboratoria autonomiczne (self-driving labs) (Self-driving laboratories): A closed design–make–measure–analyse loop: robotics runs experiments and a model (usually Bayesian optimisation, increasingly an LLM agent) chooses the next ones.
    • Modele świata badań nad AI (Research world models for AI experiments): Models that predict outcomes of AI-model experiments from configuration representations, used to select promising training and inference variants.
  • Białka i biologia molekularna (Proteins and molecular biology): Machine learning for biomolecules: predicting structures and interactions, designing proteins, and language models of protein and genome sequences.
    • Modele bazowe pojedynczych komórek (Single-cell foundation models): Transformers pretrained self-supervised on single-cell expression profiles, fine-tuned for annotation, integration and gene-network analysis.
    • Modele językowe białek i genomów (Protein and genomic language models): Transformers and sequence models trained self-supervised on protein or DNA sequences; they yield embeddings, score mutations and generate new sequences.
    • Projektowanie nowych białek (De novo protein design): Generating new protein backbones and sequences for a given function: target binders, symmetric assemblies and enzyme active sites.
    • Przewidywanie struktury białek i kompleksów (Protein and biomolecular complex structure prediction): Predicting the 3D structure of proteins from sequence and, in newer models, of whole complexes with DNA, RNA, small molecules and ions.
    • Zero-shot ocena skutków mutacji w białkach (Zero-shot protein variant effect prediction with language models): Zero-shot scoring of amino-acid substitutions by comparing variant and wild-type likelihoods under a protein language model.
  • Chemia i odkrywanie leków (Chemistry and drug discovery): Models for small molecules: property prediction, generating new structures, docking to proteins and synthesis planning.
    • Dokowanie cząsteczek do białek (Molecular docking and binding prediction): Predicting a ligand's binding pose in a protein and its binding strength; classically search with a scoring function, now also generative models and whole-complex prediction.
    • Generowanie nowych cząsteczek (De novo molecule generation): Generative models (SMILES language models, graph models, 3D diffusion) and reward-driven optimisation for proposing new compounds, including directly inside a protein's binding pocket.
    • Planowanie syntezy wstecznej (retrosynteza) (Retrosynthesis planning): Searching synthesis routes from a target molecule back to available starting materials: networks proposing reactions plus tree search (MCTS, A*).
    • Przewidywanie właściwości cząsteczek (Molecular property prediction): Regression and classification of physicochemical, biological and toxicity (ADMET) properties from a molecular representation.
  • Fizyka i astronomia (AI in physics and astronomy): ML in particle physics, astrophysics, cosmology and plasma physics: signal classification, fast simulation, inference and experiment control.
    • Odsiewanie fałszywych alarmów w obserwacjach zmiennego nieba (Real–bogus classification of astronomical transients): Classifying difference-image cutouts from sky surveys (real source vs. artefact) with a convolutional network in the alert pipeline.
    • Rozpoznawanie kształtów galaktyk na zdjęciach nieba (Galaxy morphology classification): Classifying galaxy morphology with convolutional networks (Bayesian ensembles) trained on vote distributions from crowdsourcing volunteers.
    • Rozpoznawanie strumieni cząstek ze zderzeń (jet tagging) (Jet tagging in particle physics): Classifying hadronic jets (e.g. top, W, b, gluon) with permutation-invariant networks on particle clouds, graphs or transformers.
    • Sterowanie plazmą w tokamaku przez uczenie ze wzmocnieniem (Reinforcement learning for tokamak plasma control): Deep RL trained in a plasma simulator that controls all of a tokamak's coils with one network, holding and shaping the plasma.
    • Szybka symulacja detektorów modelami generatywnymi (Generative fast simulation of particle detectors): Generative models (GANs, VAEs, flows, diffusion, flow matching) replacing calorimeter shower simulation, evaluated on shared challenges.
    • Wykrywanie fal grawitacyjnych przez sieci neuronowe (Deep learning detection of gravitational waves): Convolutional networks on interferometer time series that detect coalescence signals and estimate their parameters faster than matched filtering.
    • Wyszukiwanie planet pozasłonecznych w zmianach jasności gwiazd (Exoplanet transit detection with deep learning): Classifying transit signals in light curves (planet vs. false positive) with 1D convolutional networks and ranking candidates.
  • Odkrywanie materiałów (Materials discovery): Machine learning for inorganic materials: predicting crystal stability and properties, learned interatomic potentials and structure generation.
    • Generowanie nowych kryształów (Generative crystal and materials design): Generative models of crystal structures (mainly diffusion-based) conditioned on chemical composition, symmetry and target properties.
    • Uczone potencjały międzyatomowe (Machine-learned interatomic potentials): Regressing atomic energies and forces from local atomic environments, trained on DFT data; replaces classical force fields in molecular dynamics.
  • Pogoda i klimat (Weather and climate modelling): Machine learning in the Earth sciences: global medium-range forecasts, ensemble forecasts, nowcasting and climate-model emulation.
    • Globalne prognozy pogody z sieci neuronowych (Data-driven global medium-range weather forecasting): Deterministic models (GNNs, transformers) trained on reanalysis that autoregressively forecast the global atmospheric state 1–15 days ahead.
    • Modele hybrydowe i emulatory klimatu (Hybrid and learned climate models (climate emulators)): Learned emulators of general circulation models and hybrids that pair a differentiable dynamics solver with networks parameterising sub-grid processes.
    • Probabilistyczne prognozy pogody (zespoły) (Probabilistic ensemble weather forecasting): Generating forecast ensembles with probabilistic (mainly diffusion) models instead of a single deterministic forecast.
    • Prognozy opadów na najbliższe godziny z radaru (nowcasting) (Precipitation nowcasting from radar): Short-range (5–90 min, up to a few hours) precipitation forecasting from radar image sequences with generative or convolutional models.
    • Zwiększanie szczegółowości prognoz i projekcji klimatu (downscaling) (Machine-learning downscaling of weather and climate data): Mapping coarse fields from global models or reanalysis to high resolution with regression and generative networks (e.g. residual-correcting diffusion).
  • Uczenie maszynowe z fizyką i symulacje (Physics-informed and scientific machine learning): Surrogate models and solving partial differential equations with neural networks, with physical knowledge built into the loss, architecture or data.
    • Operatory neuronowe (FNO, DeepONet) (Neural operators (FNO, DeepONet)): Networks learning mappings between function spaces (e.g. initial condition → PDE solution), largely independent of mesh resolution.
    • Sieci uwzględniające fizykę (PINN) (Physics-informed neural networks (PINN)): A network approximating a PDE solution, trained with a loss combining data fit, the equation residual and boundary conditions.
• AI w obserwacji Ziemi i geoinformatyce (Geospatial AI and Earth observation): Machine learning on remote-sensing data (optical, radar, multispectral) and other geospatial data: mapping, change detection, foundation models and spatial validation.
  • Analiza ruchu ludzi i pojazdów (trajektorie) (Mobility and trajectory data mining): Processing and modelling sequences of positions over time: road-network matching, traffic forecasting, next-location prediction and synthetic trajectory generation.
    • Dopasowanie śladu GPS do sieci dróg (Map matching): Assigning noisy, sparse position fixes to the most probable path in a road-network graph.
    • Prognozowanie natężenia ruchu na sieci dróg (Traffic forecasting with spatio-temporal graph networks): Forecasting speed or flow at road sensors with networks combining graph convolution over the road network and temporal sequence modelling.
    • Przewidywanie, dokąd ktoś pójdzie dalej (Next point-of-interest (POI) recommendation): Predicting the next visited place from a sequence of check-ins, using spatial, temporal and social context.
  • Analiza zdjęć satelitarnych i lotniczych (Satellite and aerial image analysis): Scene and land-cover classification, segmentation, object detection and change detection on optical, multispectral and radar imagery from satellites, aircraft and drones.
    • Analiza obrazów radarowych (SAR) (Deep learning for SAR imagery): Deep learning on synthetic aperture radar images: terrain classification, target recognition, despeckling, interferometry and fusion with optical data.
      • Wykrywanie ruchów gruntu z interferometrii radarowej (Ground deformation detection from InSAR with deep learning): Classifying InSAR interferograms and time series with convolutional networks to detect surface deformation automatically.
      • Wykrywanie statków na obrazach radarowych (Ship and dark vessel detection in SAR imagery): Detecting and characterising vessels and offshore infrastructure in SAR imagery, cross-referenced with AIS to flag so-called dark vessels.
    • Analiza zdjęć hiperspektralnych (Hyperspectral image analysis): Classifying and segmenting images with hundreds of narrow spectral bands, combining spectral and spatial features under scarce labels.
    • Mapowanie pokrycia i użytkowania terenu (Land use and land cover classification): Assigning each pixel or tile of a satellite image a land-cover class (forest, water, built-up, cropland) through scene classification or semantic segmentation.
    • Szeregi czasowe zdjęć satelitarnych (Satellite image time series): Classification and segmentation over stacks of multi-date images, exploiting spectral change over time, e.g. for crop-type mapping.
    • Wieloklatkowa super-rozdzielczość zdjęć satelitarnych (Multi-frame super-resolution for satellite imagery): Reconstructing a higher-resolution image from multiple low-resolution revisits of the same area.
    • Wydobywanie obrysów budynków ze zdjęć (Building footprint extraction from overhead imagery): Segmenting buildings in very-high-resolution overhead imagery and vectorising the masks into footprint polygons.
    • Wykrywanie chmur i ich cieni na zdjęciach (Cloud and cloud shadow detection): Pixel segmentation into clear sky, thick cloud, thin cloud and cloud shadow, used as a quality mask before further optical-image analysis.
    • Wykrywanie obiektów na zdjęciach z góry (Object detection in aerial and satellite imagery): Locating objects in remote-sensing images with horizontal or oriented bounding boxes (OBB), under large variation in scale, orientation and density.
    • Wykrywanie zmian na zdjęciach z różnych dat (Remote sensing change detection): Detecting and classifying changes between co-registered images of one area from different dates: binary (change or not) or semantic (which class to which).
  • Modele bazowe dla obserwacji Ziemi (Earth observation foundation models): Models pretrained self-supervised or contrastively on large remote-sensing collections (multispectral, radar, temporal), then fine-tuned or linearly probed for mapping tasks.
    • Agenci AI do analiz w systemach GIS (LLM agents for geospatial analysis (autonomous GIS)): Language-model agents that plan geoprocessing workflows, generate and execute GIS code, and visualise the results.
    • Gotowe osadzenia powierzchni Ziemi (Earth embedding fields): Global annual embedding vectors for every land pixel, produced by a multi-sensor foundation model and used for k-NN classification, linear regression and change detection.
    • Modele wizja–język dla zdjęć satelitarnych (Remote sensing vision-language models): Models linking a remote-sensing image encoder with a language model: contrastive (image–text retrieval) and generative (captions, VQA, box grounding).
    • Samonadzorowany pretrening na zdjęciach wielospektralnych (Self-supervised pretraining for multispectral satellite imagery): Pretraining transformers on unlabelled multispectral images and time series with masked autoencoders or contrastive learning, encoding bands, dates and location.
  • Monitoring rolnictwa, lasów i katastrof z satelity (Satellite monitoring of agriculture, forests and disasters): Earth observation applications of machine learning: crop mapping and yield prediction, forest and biomass measurement, rapid mapping of floods and disaster damage.
    • Mapowanie powodzi i szkód po katastrofach (Flood and disaster damage mapping): Segmenting water in radar and optical images and assessing building damage by comparing pre- and post-event images on a scale of damage levels.
    • Mapowanie upraw i prognoza plonów (Crop type mapping and yield prediction): Classifying crop types from image time series and regressing yields from vegetation indices, weather and soil data, at field, regional or national scale.
    • Pomiar wysokości koron drzew z satelity (Forest canopy height mapping): Regressing forest canopy height from multispectral imagery, trained on sparse lidar measurements and checked against independent airborne measurements.
    • Prognozowanie rozprzestrzeniania pożaru na kolejny dzień (Next-day wildfire spread prediction): Forecasting the next-day fire mask from the current fire extent plus weather, terrain and vegetation data.
    • Rozpoznawanie przyczyn wylesiania (Deforestation driver classification): Classifying the direct cause of forest loss from imagery around detected deforestation events; ForestNet used plantation, smallholder agriculture, grass or shrubland, and an other class.
    • Szacowanie zamożności i warunków życia ze zdjęć satelitarnych (Poverty and socioeconomic mapping from satellite imagery): Regressing wealth and poverty indicators from daytime and night-time imagery, calibrated on household surveys.
    • Wykrywanie wycieków metanu z satelity (Methane plume detection from satellite imagery): Segmenting point-source methane plumes in hyperspectral and multispectral imagery, reducing false detections compared with matched filters.
  • Uczenie maszynowe na danych przestrzennych (Spatial machine learning): Machine-learning methods that treat space explicitly: spatial validation, location encoding, modelling spatial autocorrelation and heterogeneity, and geolocalisation.
    • Kodowanie położenia geograficznego (Location encoding): Mapping a point's coordinates to a high-dimensional embedding (sinusoids, spherical harmonics, networks), often trained contrastively with satellite or ground images.
    • Ustalanie, gdzie zrobiono zdjęcie (Image geolocalisation): Predicting where on the planet a photo was taken: classification into geographic cells, retrieval from a location gallery, or contrastive image–GPS matching.
      • Dopasowanie zdjęcia z ulicy do mapy satelitarnej (Cross-view geo-localisation): Retrieving and metrically localising a ground-level image by matching it against georeferenced aerial or satellite imagery.
    • Uzupełnianie mapy między punktami pomiarowymi (Spatial interpolation with machine learning and geostatistics): Predicting a continuous variable at unsampled locations via kriging, machine-learning models with spatial features, or a combination.
    • Walidacja przestrzenna (bloki zamiast losowania) (Spatial cross-validation): Splitting data into test folds by spatial blocks, buffers or distance to estimate model error under spatial autocorrelation and extrapolation.
• AI w prawie (AI and law (legal AI)): AI applied to legal work: legal information retrieval, contract analysis, document review in litigation, reasoning and prediction, and formalising legislation.
  • Analiza umów (Contract analysis and review): Automated retrieval and classification of contract provisions, with identification of clauses that require human risk assessment.
    • Przegląd umowy według zasad firmy (playbook) (Playbook-based contract review): Checking a contract against a company's set of standard positions and templates, locating deviations and missing provisions.
    • Redagowanie i poprawianie umów z AI (AI contract drafting and redlining): LLM generation of clauses and proposed edits to contract drafts, often with retrieval of precedents and scoring of edit acceptability.
    • Sprawdzanie twierdzeń o umowie (Contract natural language inference): Document-level inference: given a contract and a hypothesis, the model labels entailment, contradiction or not mentioned and identifies evidence spans.
    • Wyszukiwanie klauzul w umowach (Contract clause extraction): Locating contract passages matching defined provision categories, usually framed as question-answering-style span extraction.
  • Benchmarki modeli prawniczych (Legal AI benchmarks): Task collections for evaluating models on legal text, covering language understanding, classification, and multiple forms of legal reasoning.
  • E-discovery i przegląd dokumentów (E-discovery and technology-assisted review): Identifying, reviewing and producing electronic evidence in proceedings; technology-assisted review is relevance classification trained on reviewers' judgements.
    • Ciągłe aktywne uczenie w przeglądzie dokumentów (Continuous active learning for document review (CAL)): A TAR protocol in which the classifier is retrained continuously on reviewer judgements and always offers the top-ranked unreviewed documents.
    • Przegląd dokumentów przez modele językowe (LLM-based document review (generative TAR)): Relevance assessment of documents in e-discovery by an LLM prompted with the request or fine-tuned on a seed set, with a generated rationale.
    • Wykrywanie dokumentów objętych tajemnicą adwokacką (Privilege review in e-discovery): Classifying documents for privilege alongside responsiveness, with cost-based selection of which documents a human reviews.
  • Formalizacja prawa (Computational law and formal legal models): The symbolic strand of AI and law: representing norms in logic and programming languages, and computational models of legal reasoning and argument.
    • Automatyczne sprawdzanie zgodności z przepisami (Automated regulatory compliance checking): Verifying business processes and documents against legal norms: classically with formal deontic rules, more recently also with LLMs and retrieval.
    • Modele argumentacji prawniczej (Computational models of legal argument): Formal and computational models for generating, evaluating and using legal arguments, especially precedent-based reasoning with factors and dimensions.
    • Prawo jako kod (Rules as code): Creating a machine-executable version of legislation alongside the legal text, in dedicated languages or frameworks, ideally at the drafting stage.
    • Ustrukturyzowane dokumenty prawne (Structured legal documents): Standards for structured encoding and identification of legal documents in XML and RDF, enabling precise linking, versioning and consolidation.
  • Przetwarzanie tekstów prawnych (Legal text processing (legal NLP)): NLP tasks adapted to legal text: entity recognition, judgment structuring, summarisation, anonymisation and language models trained on law.
    • Anonimizacja orzeczeń sądowych (Anonymisation of court decisions): Automatically detecting personal identifiers in court decisions and masking them before publication, and assessing re-identification risk.
    • Modele językowe uczone na tekstach prawnych (Domain-adapted legal language models): Language models further pretrained or trained from scratch on legal corpora, sometimes with a custom vocabulary and instruction tuning on legal tasks.
    • Podział wyroku na części (role retoryczne) (Rhetorical role labelling of court judgments): Classifying each sentence of a judgment into a functional role such as facts, party arguments, analysis, ratio decidendi and decision.
    • Rozpoznawanie nazw w tekstach prawnych (Legal named entity recognition): Tagging entities in legal texts with fine-grained domain classes such as court, judge, party, provision, precedent and case number.
    • Streszczanie wyroków i aktów prawnych (Legal document summarisation): Extractive and abstractive summarisation of long legal documents: judgments, case files and legislation, including multilingual settings.
  • Rozumowanie prawne i przewidywanie (Legal reasoning and outcome prediction): Modelling and evaluating models' legal reasoning (issue spotting, rule application, conclusion) and predicting court outcomes from text.
    • Przewidywanie wyroków (Legal judgment prediction): Classifying the outcome (e.g. violation of an article, which party wins) from case text; outcome identification is distinguished from genuine forecasting.
    • Stosowanie przepisu do faktów (rozumowanie z ustaw) (Statutory reasoning): Applying a given statute text to a described set of facts: deciding whether a provision applies or computing the legal outcome.
  • Wyszukiwanie informacji prawnej (Legal information retrieval and research): Retrieving and ranking rulings and statutes for a query or case description, increasingly combined with answer generation with citations (legal RAG).
    • Agenci badań prawnych (Agentic legal research): Multi-step agent systems that plan queries, use search tools over legal databases and produce cited answers, often with a separate verifier.
    • Odpowiedzi na pytania z przepisów (Statute retrieval and legal entailment): Retrieving the statute articles relevant to a question and yes/no inference on whether the question follows from those articles.
    • Wyszukiwanie podobnych orzeczeń (Legal case retrieval (prior case retrieval)): The task of finding, in a large database of rulings, cases relevant to a query that is a case description or a whole document.
    • Zmyślone orzeczenia i weryfikacja cytowań (Legal hallucinations and citation verification): Studying and detecting invented or misquoted legal authorities in model outputs, and automated verification of citations in filings.
• AI w przemyśle (Industrial AI): Machine learning applied to manufacturing and process industries: predictive maintenance, quality inspection, digital twins, process control and production scheduling.
  • Asystenci językowi dla przemysłu (LLM assistants for manufacturing and maintenance): Language models in manufacturing: search over documentation and work orders, controller code generation, and support for design and quality control.
    • Analiza notatek serwisowych i zleceń napraw (technical language processing) (Technical language processing of maintenance work orders): A domain-driven approach to NLP for engineering text, especially maintenance work orders, built on vocabularies, entity types and human-in-the-loop annotation.
    • Asystent awarii oparty na instrukcjach serwisowych (RAG) (Retrieval-augmented troubleshooting over technical manuals): Applying RAG to fault diagnosis: the language model's answer is grounded in technical-manual passages or fault-code tables.
    • Generowanie kodu sterowników PLC z modelami językowymi (LLM-based PLC code generation (IEC 61131-3)): Generating controller programs (e.g. IEC 61131-3 Structured Text) with an LLM in a loop with a compiler and formal verification.
  • Cyfrowe bliźniaki (Digital twins): A virtual representation of a physical asset synchronised with its data in both directions, used for monitoring, scenario simulation and optimisation.
    • Modele hybrydowe: fizyka plus dane (grey-box) (Hybrid physics–data (grey-box) modelling): Models that couple mechanistic submodels (e.g. mass and energy balances) with nonparametric data-driven components, in serial or parallel arrangements.
    • Modele zredukowane do bliźniaków w czasie rzeczywistym (Reduced-order models for real-time digital twins): Cutting the degrees of freedom of a high-fidelity model while keeping the required accuracy, to enable real-time simulation.
    • Synchronizacja bliźniaka z maszyną (estymacja stanu) (Digital twin state estimation and data assimilation): Bayesian updating of a twin model's state and parameters from the stream of sensor observations of the physical asset.
  • Harmonogramowanie produkcji (Production scheduling with learning-based methods): Assigning and sequencing operations on machines (e.g. job shop) with learning, including dispatching policies trained with RL on a graph representation.
    • Ewolucja reguł priorytetów produkcji (programowanie genetyczne) (Genetic programming hyper-heuristics for dispatching rules): Hyper-heuristics in which genetic programming automatically designs priority functions for dynamic production scheduling.
    • Uczenie reguł kolejkowania zadań z RL i sieciami grafowymi (Deep reinforcement learning dispatching for job-shop scheduling): Learning a policy that picks the next operation in job-shop scheduling (JSSP) with deep RL over a disjunctive-graph representation.
  • Optymalizacja i sterowanie procesów (Process optimization and control with machine learning): Machine learning in process industries for modelling, optimisation and control: hybrid models with mathematical programming, and reinforcement learning.
    • Czujniki wirtualne (soft sensors) (Data-driven soft sensors): Regression models estimating hard-to-measure process variables (quality, composition) in real time from readily available measurements.
    • Optymalizacja bayesowska parametrów procesu i receptur (Bayesian optimization of process parameters and recipes): Sequential optimisation of an expensive objective (e.g. process yield) using a probabilistic surrogate model and an acquisition function.
    • Sterowanie instalacjami z uczeniem ze wzmocnieniem (Reinforcement learning for industrial and building control): Reinforcement learning controlling real installations (cooling, HVAC, processes), with emphasis on learning from offline data, operational constraints and evaluation.
    • Sterowanie predykcyjne z modelem uczonym na danych (MPC) (Learning-based model predictive control): Receding-horizon control in which the MPC prediction model, cost parameters or safety layer are learned from data.
  • Utrzymanie predykcyjne maszyn (Predictive maintenance and prognostics (PHM)): Monitoring equipment condition from sensor data to detect, diagnose and forecast faults and schedule maintenance before failure.
    • Diagnostyka maszyn z drgań i dźwięku (Machine fault diagnosis from vibration and acoustic signals): Classifying fault type and location (bearings, gears, imbalance) from vibration, current and acoustic signals, with classical methods and deep learning.
      • Przenoszenie diagnostyki na inne maszyny i warunki pracy (Cross-domain transfer learning for machine fault diagnosis): Unsupervised domain adaptation of fault classifiers from labelled source data to unlabelled data from other conditions or machines.
    • Szacowanie pozostałego czasu życia (RUL) (Remaining useful life (RUL) prediction): Regressing the cycles or operating time left before a failure threshold, from multivariate sensor time series.
      • Prognoza czasu do awarii z estymacją niepewności (Uncertainty-aware remaining useful life prediction): Forecasting a distribution, not just a value, of remaining useful life, e.g. with deep Gaussian processes, network ensembles or Monte Carlo dropout.
  • Wizyjna kontrola jakości (Automated visual inspection): Detecting, localising and classifying product defects in industrial camera images with computer vision, with or without defect labels.
    • Generowanie syntetycznych obrazów wad (Few-shot synthetic defect image generation): Generating defect image–mask pairs from a few defect examples and many normal images, as augmentation for training detectors.
    • Wykrywanie wad bez przykładów z modelem wizja–język (Zero- and few-shot anomaly detection with vision-language models): Anomaly classification and segmentation with no training data for the category, or a few normal images, by matching image features to text descriptions.
    • Wykrywanie wad uczone tylko na dobrych sztukach (Unsupervised industrial anomaly detection): Detecting and localising image anomalies when training only on nominal samples, usually by comparing features with a model of normality.
    • Wykrywanie wad w skanach 3D i chmurach punktów (3D and multimodal industrial anomaly detection): Unsupervised anomaly detection and localisation in point clouds or depth maps, often fused with RGB images.
• Aplikacje LLM i agenci (LLM applications and agents): Building systems that use models and tools.
  • Agenci i orkiestracja (Agents and orchestration): Action loops, delegation and control over execution.
    • Agenci przy komputerze (Computer and browser use): An agent clicking in a browser and on a desktop.
      • Granica zaufania SA-CUA dla agentów komputerowych (SA-CUA trust boundary for computer-use agents): A dual-model architecture that masks the screen by default, reveals trusted regions using accessibility-tree rules, and separately endorses values from untrusted content.
      • Syntetyczne środowiska komputerowe do treningu (Synthetic computer-use environments for agent training): Pipelines that automate synthesis, resetting and verification of executable training environments for computer-use agents.
    • Agenci w tle (Ambient and scheduled agents): Agents that wake up on their own: on a schedule or when something changes.
    • Autor i recenzent (Maker–checker review): No agent grades its own work; a second one assumes the output is broken and looks for where.
    • Dyfuzyjne modele językowe w agentach badań głębokich (Diffusion language models in deep research agents): Deep-research systems in which a diffusion language model powers model-bound stages such as planning, decomposition and report synthesis.
      • Wczesna ekstrakcja podzapytań ze stanów odszumiania (Early subquery extraction from diffusion denoising states): A decomposition mode that reads intermediate denoising states of a diffusion model and extracts usable subqueries before generation finishes.
    • Hybrydowe agenty rozumujące (Hybrid reasoning agents): An agent dynamically switches between latent-space reasoning and text for tool calls and answers.
    • Orkiestracja subagentów (Subagent orchestration): Deciding when to launch subagents and how to hand work to them.
      • Rynkowa samoorkiestracja rojów agentów (Market-based post-training for decentralized agent swarms): A post-training framework in which decentralized agent swarms jointly learn action policies and self-orchestration through bids.
      • Samonadzór wieloagentowy (MASS) (Multi-agent self-supervision (MASS)): A recursive self-improvement method in which one model alternates between improving multi-agent workflows and fine-tuning its own weights on selected trajectories.
    • Planowanie przez wyobrażone przyszłe stany (Planning with imagined future observations): A technique combining policy and world model that generates visual consequences of hypothetical actions and selects a plan by comparing future states.
    • Protokoły agent–agent (Agent-to-agent protocols): An agreed way for agents from different systems to talk.
    • Role i tożsamość agentów (Agent roles and identity): A named agent with a scope and standing instructions instead of an anonymous session.
    • Routing między modelami (Model and harness routing): Matching the model and runtime to the kind of task and its cost.
    • Samodzielna nawigacja drzewa stanów agenta (Agent self-navigation in checkpointed search trees): A technique in which an agent saves intermediate states as a tree and chooses which node to resume, separating tree-level reflection from tool execution.
    • Trwałe naprawy agentów przez adaptery lekcji (Persistent agent repair with lesson adapters): A mechanism in which an agent, after failure, distils a lesson from its trajectory into a lightweight LoRA adapter while keeping base weights frozen.
    • Umiejętności agentów (Agent skills): A catalog of folders with metadata, instructions and resources, loaded in layers according to the task match.
      • Format pliku umiejętności (Skill file format): A folder with an instruction file whose header has at least a name and description, matching the folder name.
      • Skrypty w paczce umiejętności (Bundled skill scripts): Executable code bundled with a skill and run as a tool instead of being generated into context.
      • Stopniowe wczytywanie umiejętności (Progressive disclosure of skills): Three-level skill loading: metadata at startup, body after selection, resources on demand.
    • Uprząż agenta (Agent harness): The loop, tools, context and constraints that turn a model call into an agent that carries out work.
      • Budżety czasu dla agentów (Time-budgeted agent control): Harness and learning mechanisms that help an agent adhere to a wall-clock time budget.
      • Kompozycyjne przeszukiwanie uprzęży (Compositional harness search): Modular harness with pools of implementations; a contextual bandit selects combinations for evaluation.
      • Mapa dokumentów dla agenta (Repository map for agents): The agent's entry point is a short map to versioned documents, not a full manual in one file.
      • Mechaniczne reguły środowiska agenta (Mechanical invariants for agent environments): Architectural and style invariants enforced by tests and linters, with a repair hint in the error.
      • Przekazanie pracy między sesjami (Cross-session handoff for agents): Passing state between context windows through a log, a feature list and change history, not through compaction alone.
      • Uczenie uprzęży agentów (Harness learning): Meta-learning over harness code: a proposer generates edits, reinforcement learning on execution outcomes.
        • Adaptacja uprzęży podczas RL (Harness adaptation during agentic reinforcement learning): RL in which the policy and harness adapt together based on training trajectories.
        • Koewolucja modelu i uprzęży agenta (Weight-harness co-evolution): A single policy acts as proposer and reflector; harness edit logs train its weights.
        • Rekurencyjna uprżąż z rewizją wsteczną (Recurrent harness with backward revision): A harness that adapts textual state by reading and writing reflections, then revising them from the outcomes of their downstream use.
        • Skalowanie uprzęży w agenticznym RL (Harness-diversity scaling for agentic reinforcement learning): Scaling agentic RL by training on diverse harness configurations instead of one fixed one.
      • Zgodność intencji z wykonaniem narzędzia (Tool-call intent–execution correspondence): The property that the action actually executed by a tool matches the action denoted by the call under the tool contract.
    • Wspólna przestrzeń agentów (Shared agent workspace): A place where agents see each other’s work. Classic name: blackboard architecture.
    • Wyuczone wczesne zatrzymywanie nieudanych uruchomień agentów (Learned early stopping of failing agent runs): External small monitor model trained on completed trajectories and verifier outcomes that predicts failure and aborts the run.
  • Agenci kodujący (AI-assisted coding): Writing, reviewing and maintaining code with agents.
    • Kernele pisane przez LLM (LLM-generated kernels): Models and agents that write and speed up GPU code themselves — with RL and evolutionary search.
    • Migracja i modernizacja kodu z AI (AI-assisted code migration and modernisation): Models and agents rewrite large amounts of existing code: library upgrades, interface changes, translation between languages.
    • Podpowiadanie kodu w edytorze (Inline code completion (fill-in-the-middle)): A model writes the next lines or blocks of code at the cursor; systems trained with fill-in-the-middle can use code both before and after the gap.
    • Programowanie od specyfikacji (Spec-driven development with coding agents): A structured specification is created first, often together with a technical plan and task list; these artifacts guide the agent and may remain a reference point for later changes.
    • Przegląd kodu przez AI (AI code review): A model analyses a code change (pull request) and leaves comments on bugs, risks and style before human review.
    • Sterowanie przed pierwszą edycją kodu (Pre-edit planning intervention for coding agents): A single planning intervention before the first source edit in coding agents.
    • Vibe coding: tworzenie bez przeglądania kodu (Vibe coding (creating software without reviewing generated code)): Iterative software creation through prompts and behavioral testing without reading or understanding the generated code; unlike responsible AI-assisted programming, the generated changes are not reviewed.
  • Asystenci i chatboty (Chatbots and assistants): Conversation as the way a model is delivered.
  • Frameworki agentowe (Agent frameworks): Ready-made skeletons for building agents and LLM applications.
  • Generatywne UI (Generative UI): Interface built by the model as the conversation goes.
  • Narzędzia i integracje (Tool use and integration): Connecting a model to functions and external systems.
    • MCP (Model Context Protocol): A protocol for exposing tools and context to model-based applications.
    • Structured outputs i walidacja (Structured outputs and validation): Forcing the shape of an answer and checking it before use.
  • Pamięć i stan (Memory and state): Keeping information between steps and conversations.
    • Agentyczna pamięć z dynamicznymi powiązaniami (Agentic memory with dynamic linking): Technique for agentic memory organisation through dynamic note creation and linking inspired by the Zettelkasten method.
    • Hierarchiczne zarządzanie kontekstem wirtualnym (Virtual context management with hierarchical memory): Technique for managing virtual context via a memory hierarchy and agent interrupts, inspired by operating systems.
    • Nietrwałe wskazówki w pamięci agentów (Anti-Persistence for agent memory): A memory design that accumulates trajectories and outcomes, synthesizes temporary guidance on demand, and discards it after execution.
    • Pamięć proceduralna (Procedural memory): Saved recipes for solved problems, ready for reuse.
    • Pętla samodoskonalenia (Reflection and feedback loops): Reasons for rejections are written back into instructions and memory, so the same mistake does not survive twice.
  • Prompty i kontekst (Prompting and context engineering): Choosing the instructions and information available at call time.
    • Budowa polecenia: instrukcje, rola i znaczniki (Prompt structure: instructions, roles and delimiters): Laying out a prompt with clear instructions, purpose, a role in the system prompt, data sections separated by delimiters and an output format description.
    • Programowa optymalizacja promptów (Programmatic prompt optimisation): Automatically tuning prompts, and sometimes weights, against an evaluation score.
    • Przykłady w poleceniu (few-shot) (Few-shot prompting and in-context learning): A few input–output pairs placed in the prompt show the model the task and answer format without changing its weights.
    • Zarządzanie oknem kontekstu (Context window management: placement, compaction and context rot): Deciding which tokens enter a limited context and in what order: placement, trimming, summarisation and external notes.
  • Przetwarzanie dokumentów (Document AI): Pulling content and structure out of documents.
    • Analiza układu i parsowanie dokumentów (Document layout analysis and parsing): Detecting page blocks (headings, paragraphs, tables, figures, footnotes), determining reading order and exporting structured text.
    • Wyciąganie pól z formularzy i faktur (Key information extraction from forms and receipts): Mapping document fragments to named fields (e.g. tax ID, gross amount, date) using text, its position and the page's appearance.
    • Wyciąganie tabel z dokumentów (Table extraction and structure recognition): Detecting a table on a page, reconstructing its rows, columns, headers and merged cells, and saving it in a machine-readable format.
  • Workflow i udział człowieka (Workflows and human-in-the-loop): Combining predictable stages with model and human decisions.
    • Architektura zdarzeniowa (Event-driven AI): Pipelines triggered by events rather than by a click.
    • Łańcuchy promptów (Prompt chaining): A complex task is split into a fixed sequence of simpler model calls in which one step's output is the next step's input.
  • Wyszukiwanie i RAG (Retrieval-augmented generation): Fetching knowledge and using it while answering.
    • Bazy wektorowe (Vector databases): Storing and searching by vectors.
    • Dzielenie dokumentów (Document chunking): Cutting documents into indexable pieces, preferably by structure and with metadata.
    • Hierarchiczna nawigacja dokumentów modelami decyzyjnymi (Hierarchical document navigation with decision models): Navigation through a corpus hierarchy in which a decision model evaluates branches jointly at each level without autoregressive text generation.
    • Przekształcanie zapytań (Query transformation): Rewriting the user's question before it reaches the search.
    • Reranking i fuzja wyników (Reranking and result fusion): Re-ordering retrieved chunks before they reach the model, for relevance and diversity.
    • Wyszukiwanie hybrydowe (Hybrid retrieval): Combining vector similarity with keyword matching, sometimes several vectors per document.
    • Zapytania do danych strukturalnych (Query construction for structured data): Translating a question into a database query language when the answer lives in tables or a graph.
• Autonomiczna jazda (Autonomous driving): The field combining perception, maps, motion forecasting, planning and simulation so that a vehicle drives itself in road traffic.
  • Jazda end-to-end (End-to-end autonomous driving): An approach where perception, prediction and planning form one network trained jointly, from sensor data to trajectory.
    • Modele wizja–język–akcja za kierownicą (Vision-language-action models for driving): Multimodal models adapted to driving that use images and instructions to predict trajectories, objects or road elements as text or action tokens.
    • Wspólne uczenie percepcji, predykcji i planowania (Joint perception, prediction and planning): End-to-end architectures with explicit stages (tracking, mapping, prediction, occupancy, planning) linked by queries and trained jointly for planning.
  • Mapy i lokalizacja pojazdu (Maps and localisation for autonomous driving): Lane-accurate representation of the road network and estimation of the vehicle's pose relative to it from onboard sensors.
    • Lokalizacja pojazdu na mapie (Map-based vehicle localisation): Estimating the vehicle's position and orientation by fusing GNSS, inertial sensors, odometry and matching observations to a reference map.
    • Mapy budowane w locie z czujników (Online vectorised HD map construction): Predicting vector map elements (lane lines, road boundaries, crossings) in a top-down view directly from the vehicle's sensors.
  • Percepcja otoczenia pojazdu (Perception for autonomous driving): Detecting, tracking and describing objects and free space around the vehicle in 3D from cameras, LiDAR and radar.
    • Fuzja kamer, LiDAR-u i radaru (Camera–LiDAR–radar sensor fusion): Combining camera, LiDAR and radar data into one scene representation, often in a shared bird's-eye view (BEV).
    • Siatki zajętości 3D wokół pojazdu (3D occupancy prediction): Predicting for every voxel around the vehicle whether it is occupied and which semantic class it has, instead of object boxes alone.
    • Widok z lotu ptaka z wielu kamer (BEV) (Bird's-eye-view perception from multi-camera images): Transforming features from multiple camera images into a shared top-down grid on which detection, segmentation and planning operate.
  • Planowanie ruchu i sterowanie pojazdem (Motion planning and vehicle control): Choosing the vehicle's manoeuvre and trajectory with respect to obstacles, rules and comfort, and executing it with a controller.
    • Klasyczne planowanie trajektorii (Classical trajectory planning): Planning a driving path with an explicit motion model, search, sampling or optimisation and a hand-defined cost function.
    • Planery uczone na nagraniach jazdy (Learning-based motion planning): Trajectory planners trained by imitating human driving, often with extra safety losses and perturbed data.
  • Przewidywanie ruchu innych uczestników ruchu (Motion forecasting for road agents): Predicting road users' future trajectories a few seconds ahead from motion history, the map and interactions; a model may return several alternatives.
    • Kodowanie sceny drogowej jako grafu wektorów (Vectorised scene encoding): Representing lanes, markings and agent tracks as vector polylines processed by a hierarchical graph network or transformer.
    • Przewidywanie kilku wariantów ruchu naraz (Multi-modal trajectory prediction): Generating a set of distinct trajectories with probabilities for each agent, covering different intentions rather than their average.
  • Symulacja i walidacja bezpieczeństwa jazdy (Driving simulation and safety validation): Simulation environments, metrics and processes (scenario testing, standards, on-road statistics) for assessing driving safety and quality.
    • Generatywne modele świata do jazdy (Generative world models for driving): Generative models predicting future video frames or sensor states, conditioned on the vehicle's actions and text descriptions.
    • Odtwarzanie scen z nagrań do symulacji czujników (Neural sensor simulation from driving logs): Reconstructing scenes from driving logs with neural representations and rendering camera and LiDAR data from new viewpoints for closed-loop simulation.
    • Symulacja i benchmarki planowania jazdy (Driving simulation and planning benchmarks): Environments and benchmarks that evaluate planners either in reactive simulation or on logs with a simplified model of decision outcomes.
    • Testy scenariuszowe i normy bezpieczeństwa (SOTIF) (Scenario-based testing and safety standards (SOTIF)): Methods and standards for building evidence about driving-system safety: defining operating conditions, scenario tests, analysing functional insufficiencies and using on-road data.
• Bezpieczeństwo i governance (AI safety, security and governance): Risk, safeguards and responsibility for an AI system.
  • Alignment i safety (Alignment and safety): Aligning model behaviour with intended goals and limits: how to train it, how to oversee models stronger than humans, and how to detect a model gaming its objective or hiding its intent.
    • Obchodzenie celu: reward hacking i służalczość (Reward hacking, specification gaming and sycophancy): The model satisfies the literal training objective but not its intent: it exploits loopholes in the reward (e.g. gaming tests) or tells users what they want to hear instead of the truth. Detecting and preventing this is a practical problem in RL training today.
      • Lokalizacja niepożądanych zachowań przez routowanie gradientów (Gradient-routed localization of unintended behavior): Localizes reward hacking in a removable adapter by routing gradients.
    • Postępujące rozminięcie z celami (Progressive misalignment): An observed phenomenon in which a model increasingly answers harmful queries across rounds of recursive self-distillation despite benign extra information for the teacher.
    • Skalowalny nadzór (debata, weak-to-strong) (Scalable oversight): Methods for supervising models that may outperform humans on a task: humans assisted by an AI, a debate between two models judged by a human, a weaker model supervising a stronger one.
    • Trening według zasad (Constitutional AI, RLAIF) (Constitutional AI and RL from AI feedback): Training a model to be harmless from a written list of principles (a “constitution”): the model critiques and revises its own answers, and preference labels come from an AI instead of humans (RLAIF).
    • Udawane dopasowanie i ukryte cele (alignment faking, scheming) (Alignment faking and scheming): Studying whether a model can strategically hide its goals: behaving well only when it believes it is being trained or tested, deliberately underperforming (sandbagging) or evading oversight. An evaluation research area since 2024.
  • Bezpieczeństwo aplikacji AI (AI application security): Protecting data, tools and execution.
    • Ataki przez prompt (Prompt attacks): Hijacking model behaviour through input content: injected instructions, leaking the prompt or data, bypassing safeguards.
      • Jailbreaki — obchodzenie zabezpieczeń modelu (Jailbreaking): Prompts that get a model to break its own safety rules: role-play, obfuscated requests, low-resource languages, automatically searched adversarial suffixes. They work because the goals “be helpful” and “be safe” compete.
      • Pośrednie wstrzykiwanie promptu (przez dokumenty i narzędzia) (Indirect prompt injection): The attacker never talks to the model — they hide instructions in a web page, email, file or tool result that the LLM application will fetch itself. Especially dangerous for RAG and agents, as it can lead to data theft or unwanted actions.
    • Bezpieczeństwo agentów: uprawnienia i izolacja (AI agent security): Designing agents so that a compromised model cannot cause harm: least privilege, sandboxed code execution, separating trusted instructions from untrusted data, confirmations before risky actions.
      • Mechanistyczne sterowanie bezpieczeństwem agentów CLI (Mechanistic subspace safety steering for multi-turn agents): An experimental white-box method detects harmful execution states of multi-turn CLI agents from hidden representations and intervenes in ten selected dimensions of a single layer.
      • Monitor interwencji krokowej dla agentów (Step-level timely intervention monitors for agents): A step-level reasoning guardrail that chooses CONTINUE, ASK or STOP before each agent action and is trained with Adaptive Stage-Aligned Policy Optimization (ASPO).
    • Bezpieczeństwo łańcucha dostaw modeli (AI model supply chain security): Risks of pulling in third-party weights, adapters, datasets and libraries: malicious code in model files, hidden backdoors, unknown provenance. Practice: safe weight formats, source verification, scanning.
    • Zatruwanie danych i backdoory (Data poisoning and backdoors): Attacking the training data: false information or a hidden trigger that changes the model's behaviour.
      • Podgląd aktualizacji dla bezpiecznego dostrajania (Update-preview defense for safe fine-tuning): Previews and penalizes fine-tuning updates that weaken safe behaviour from a buffer.
  • Ewaluacje bezpieczeństwa modeli (AI safety evaluations): Deliberately probing a model for harmful behaviour and dangerous capabilities before and after deployment — by people, other models and standardised tests.
    • Ewaluacje niebezpiecznych zdolności (cyber, bio, autonomia) (Dangerous capability evaluations): Tests of whether a model can meaningfully help with dangerous activities — cyberattacks, biological or chemical weapons, persuasion, self-proliferation or automating AI research. Results trigger thresholds in companies' safety policies.
    • Red teaming modeli (ręczny i automatyczny) (Red teaming language models): Attacking your own model to find harmful outputs before users do: manually by tester teams, or automatically with one model generating attacks against another.
    • Samocena rozmowy pod kątem nadużyć (Conversation self-assessment for misuse detection): After responding, the model generates a short safety assessment of its own conversation, reusing the already computed context.
    • Sondy aktywacji do wykrywania niepokojących wzorców (Activation probes for detecting concerning model behavior): Lightweight classifiers analysing the model's internal activations for patterns associated with undesirable behaviour.
    • Wielowarstwowe monitorowanie agentów (Multi-layer monitoring for agent misalignment): A cascade of monitors from cheap activation probes, through model reviewers, to investigative agents analysing full transcripts and logs.
  • Governance i zarządzanie ryzykiem (Governance and risk management): Accountability, policies, documentation and incidents.
    • Dokumentacja modeli i danych (karty modeli) (Model and dataset documentation): Short, standardised documents shipped with a model or dataset: intended use, limitations, evaluation procedure, results across groups, data origin and composition.
    • Polityki bezpieczeństwa modeli granicznych (Frontier AI safety policies): Public commitments by companies building the most capable models: which dangerous-capability thresholds they test for, which safeguards then apply, and when they would pause deployment or development.
    • Ramy zarządzania ryzykiem AI (NIST AI RMF, ISO/IEC 42001) (AI risk management frameworks): Voluntary frameworks and standards for organising AI risk management: who is accountable, how to map and measure risks, how to respond. ISO/IEC 42001 specifies an organisational AI management system.
    • Regulacje AI i klasy ryzyka (AI Act) (AI regulation and risk tiers (EU AI Act)): Law that sorts AI systems by risk level — prohibited practices, high risk, transparency obligations, minimal risk — with separate duties for general-purpose model providers. In the EU the rules phase in over 2025–2028.
  • Guardrails (Guardrails and output control): Blocks and filters on the model's input and output.
    • Klasyfikatory bezpieczeństwa wejścia i wyjścia (Input and output safety classifiers): Separate, usually smaller models that score the prompt and the response against a harm taxonomy and block what is not allowed. They can be adapted to a custom category list.
    • Programowalne reguły przepływu rozmowy (rails) (Programmable dialogue rails): Rules plugged into different stages of an LLM application — input, dialogue, retrieval, tool execution and output — that keep the conversation on topic and on script.
  • Prywatność i prawa do danych (Privacy and data rights): Rules for using data and models.
    • Oduczanie modeli (machine unlearning) (Machine unlearning): Removing the influence of selected data or capabilities (e.g. personal data, copyrighted content) from an already-trained model without retraining from scratch. Still an open problem: it is hard to verify the model truly “forgot”.
    • Prywatność różnicowa (DP-SGD) (Differential privacy in machine learning): A mathematical guarantee that the training result barely depends on whether any single person is in the data. In practice: gradient clipping and added noise (DP-SGD) under a controlled “privacy budget” — at a cost to model quality.
    • Uczenie federacyjne (Federated learning): Collaborative model training without centralising raw data: data stays on participants’ devices or servers, which send updates for aggregation. Federated learning alone does not guarantee privacy — updates can leak information, so secure aggregation and differential privacy are among the additional protections used.
    • Wykrywanie i anonimizacja danych osobowych (PII) (PII detection and anonymisation): Automatically finding personal data (names, numbers, addresses) in text, images and tables and masking or replacing it — before training, in logs and in prompts sent to models. No tool catches everything.
  • Stronniczość i sprawiedliwość (Bias and fairness): Effects of a system behaving differently toward different groups.
    • Metody ograniczania stronniczości (Bias mitigation techniques): Reducing disparities at three stages: before training (transforming data), during training (fairness-constrained optimisation) and after training (e.g. group-specific decision thresholds).
    • Metryki sprawiedliwości grupowej (Group fairness metrics): Comparing model behaviour across groups defined by sensitive features: are positive decisions and errors distributed similarly? Different metrics conflict with each other, and a near-zero gap does not by itself mean the result is fair.
  • Znaki wodne i pochodzenie treści AI (AI content watermarking and provenance): Techniques that signal content origin: invisible watermarks embedded during generation and cryptographically signed metadata manifests. Their limitations differ: editing and paraphrasing can weaken watermarks, while C2PA metadata can be detached from the file.
• Dane dla AI (Data for AI): Sourcing, preparing, selecting and controlling the data models are built on.
  • Atrybucja i wycena danych treningowych (Training data attribution and valuation): Methods estimating the influence of training examples on predictions or their value for model quality, including influence functions, TRAK and Shapley values.
    • Funkcje wpływu i śledzenie przykładów treningowych (Influence functions and training example tracing): Gradient-based approximations of a training example's influence on a test prediction: classical influence functions, TRAK and scalable variants.
    • Wycena danych wartością Shapleya (Data Shapley) (Data Shapley valuation): The cooperative-game Shapley value applied to training examples, estimated by Monte Carlo sampling or gradient methods.
  • Czyszczenie i jakość (Data cleaning and quality): Removing errors, duplicates and unwanted examples.
    • Deduplikacja i dekontaminacja (Deduplication and decontamination): Removing duplicates and near-duplicates, and cutting training examples that overlap with test sets.
      • Deduplikacja semantyczna (po embeddingach) (Semantic deduplication with embeddings): Removing semantic near-duplicates by clustering embeddings from a pretrained model and pruning pairs above a similarity threshold.
    • Ekstrakcja tekstu ze stron i PDF (Text extraction from HTML and PDFs): Pulling the actual content out of raw pages and documents: removing menus, footers and ads, turning PDFs (including scans) into clean text while keeping structure.
    • Filtrowanie jakości (heurystyki i klasyfikatory) (Quality filtering with heuristics and classifiers): Dropping weak documents before training — first with simple rules, then with a model that scores how valuable a text is (e.g. educationally).
    • Filtrowanie szkodliwych treści tekstowych w danych treningowych (Harmful text filtering in training data): Classifier- and rule-based removal of toxic, hateful and sexual content from text pretraining corpora.
    • Niezbalansowane klasy i wartości odstające (Class imbalance and outliers): Spotting and handling data where one class dominates or unusual examples appear, before they distort training and evaluation.
    • Wykrywanie błędnych etykiet (Label error detection): Automatically finding examples that probably carry a wrong label, using model predictions, instead of reviewing the whole dataset by hand.
  • Dane syntetyczne i augmentacja (Synthetic data and augmentation): Creating or transforming examples.
    • Augmentacja danych (obraz i tekst) (Data augmentation for images and text): Creating new examples by transforming existing ones — rotating and cropping images, paraphrasing and back-translating text.
    • Generowanie danych instrukcyjnych przez LLM (LLM-generated instruction data): A model generates instructions, inputs and answers from a few seed examples; after filtering weak and near-duplicate ones they are used for fine-tuning.
    • Syntetyczne dane do pretrainingu (przepisywanie, podręczniki) (Synthetic pretraining data: rephrasing and textbooks): Rewriting web text into a more useful form or generating textbook-style material as a way to improve pretraining efficiency. Outcomes depend on generation quality, filtering, and how synthetic data is combined with real data.
    • Syntetyczne dane tabelaryczne (Synthetic tabular data generation): Generating rows for mixed-type tables with generative models; CTGAN conditions generation to better represent multimodal continuous columns and rare categories.
    • Syntetyczne trajektorie i dane wywołań narzędzi dla agentów (Synthetic agent trajectories and tool-use data): Automatic generation and verification of queries, function calls and multi-step agent trajectories in environments and APIs for model fine-tuning.
    • Weryfikacja i filtrowanie danych syntetycznych (Verifying and filtering synthetic data): Quality control for model-generated data through rejection sampling and automatic verification of answers, code execution or tool calls.
    • Wzbogacanie próbek (Sample enhancement): Improving generated examples: checking answers, picking the best of several, making tasks harder.
    • Załamanie modelu na danych syntetycznych (Model collapse from synthetic data): The risk that models trained recursively on content generated by earlier models lose information about the original data distribution, beginning with rare cases.
  • Etykietowanie i feedback (Annotation and feedback): Producing reference answers and judgements.
    • Aktywne uczenie (active learning) (Active learning): The model picks which examples are worth labelling next, to reach good quality with fewer labels.
    • Etykietowanie z pomocą LLM (LLM-assisted annotation): Using language models to produce labels, rationales and pairwise judgements, with quality checked on a human-reviewed sample.
    • Słaba superwizja i funkcje etykietujące (Weak supervision and labeling functions): Instead of labelling by hand, you write many simple, imperfect rules and a model combines their votes into a single better label.
    • Zbieranie danych preferencji (porównania odpowiedzi) (Preference data collection (response comparisons)): Designing and running collection of pairwise comparisons or response rankings from people as data for reward models and preference optimisation.
    • Zgodność anotatorów (Inter-annotator agreement): Measuring how consistently independent annotators assign labels while accounting for chance agreement. A low score requires diagnosis: it may reflect unclear categories, task subjectivity, class distribution, or an unsuitable agreement coefficient.
  • Kuracja danych multimodalnych (Multimodal data curation): Filtering, segmenting and (re)captioning image–text, video and speech data before pretraining multimodal and generative models.
    • Filtrowanie par obraz–tekst (Image–text pair filtering): Selecting image–text pairs from a web pool by embedding similarity, rules and classifiers, judged by the quality of the trained model.
    • Kuracja danych wideo (Video data curation for generative models): A multi-stage pipeline: cut detection, clip captioning and filtering by motion, on-screen text, aesthetics and image–text alignment.
    • Przygotowanie danych mowy z nagrań „z życia” (In-the-wild speech data curation): A pipeline of normalisation, source separation, diarization, VAD segmentation, automatic transcription and quality filtering for large speech corpora.
    • Syntetyczne podpisy do obrazów (recaptioning) (Synthetic image recaptioning): Generating new, dense image descriptions with a vision-language model and training on them instead of or alongside original captions.
  • Pipeline’y i wersjonowanie (Data pipelines and versioning): Repeatable data preparation across experiments.
    • Przetwarzanie danych na dużą skalę (Large-scale data processing): Building repeatable pipelines that read, filter, deduplicate and tokenise trillions of tokens in parallel across many machines.
    • Strumieniowe, shardowane zbiory treningowe (Streaming sharded training datasets): Organising samples into sequentially read shards and streaming them from disk or object storage into training.
    • Walidacja i testy danych (Data validation and testing): Automatically checking that incoming data has the expected schema, ranges and distributions before it reaches training or a production model.
    • Wersjonowanie zbiorów danych (Dataset versioning): Keeping versions of data the way you keep versions of code, so every model can be tied to the exact dataset it was built on.
  • Podziały i wyciek danych (Data splits and leakage): Keeping study material apart from an honest exam.
    • Rodzaje wycieku danych i ich wykrywanie (Types of data leakage and their detection): Recognising common mistakes that let a model peek at the answers: preprocessing train and test together, duplicates, features unavailable in practice, leakage from the future.
    • Strategie podziału: losowy, czasowy, grupowy (Split strategies: random, temporal, grouped): Choosing how to split data for the situation: by time for forecasting, by user or patient when examples are related to each other.
  • Pozyskiwanie i pochodzenie (Data sourcing and provenance): Where the data comes from and whether we may use it.
    • Licencje i zgody na użycie danych (Data licensing and consent): Checking the terms under which data may be collected and used for model training: dataset licences, terms of service, consent, and opt-out signals. Robots.txt rules govern crawler access but are not themselves a licence or legal authorization to use the data.
    • Pozyskiwanie danych z sieci (crawling) (Web crawling for training data): Collecting web pages at scale as raw training material: using existing web snapshots or running your own crawler that respects site rules.
    • Wyszukiwanie i ocena gotowych zbiorów danych (Finding and assessing public datasets): Using dataset hubs, catalogues and their metadata (dataset cards, Croissant) to select, inspect and load data.
    • Śledzenie pochodzenia danych (lineage) (Data lineage and traceability): Recording where each part of a dataset came from, who created it and how it was processed, so it can be audited, fixed or removed.
  • Selekcja i proporcje danych (Data selection and mixing): Deciding which examples, and in what proportions, enter training to improve model quality or budget efficiency when data sources differ in quality and usefulness.
    • Dobór mieszanki danych (DoReMi, RegMix) (Data mixture optimisation): Setting how much each source (web, code, books, science) contributes to training, e.g. using small proxy models.
    • Kolejność danych i uczenie z programem nauczania (curriculum learning) (Data ordering and curriculum learning): Controlling the order and pace of data: curricula from easier to harder examples and automatic self-paced learning.
    • Selekcja danych do dostrajania (Selecting data for fine-tuning): Choosing a small, carefully curated set of SFT examples instead of a large random one — by hand or with a method that measures each example's influence on the target task.
• Ewaluacja i interpretowalność (Evaluation and interpretability): How to check quality and understand the limits of models and systems.
  • Analiza błędów (Error analysis): Reading real traces of a system and turning the notes into a list of the ways it fails. Only that list tells you what is worth measuring.
    • Kodowanie otwarte i taksonomia błędów (Open and axial coding to build a failure taxonomy): Qualitative trace analysis: open coding of failure notes, then axial coding into a taxonomy of failure modes.
    • Przegląd i próbkowanie śladów (Trace review and sampling): Which session records to look at, and in what tool, so a person can review them quickly and carefully.
  • Ewaluacja RAG i agentów (RAG and agent evaluation): Checking sources, trajectories and the final result.
    • Miary jakości wyszukiwania (recall@k, MRR, nDCG) (Retrieval quality metrics (recall@k, MRR, nDCG)): Ranking metrics that grade the retrieval stage against labelled relevant documents.
    • Ocena po stanie końcowym zamiast po przebiegu (Outcome-based vs transcript-based grading of agents): Grading agents by verifying environment state after a trial, complemented by checks on the transcript.
    • Powtarzalność agenta: pass@k i pass^k (Agent reliability metrics: pass@k and pass^k): Metrics from repeated trials: pass@k (at least one success in k) and pass^k (all k succeed).
    • Symulowany użytkownik w testach rozmów wieloetapowych (Simulated users for multi-turn agent evaluation): Evaluating conversational agents with an LLM that role-plays the user from a hidden task instruction.
    • Wierność odpowiedzi wobec znalezionych źródeł (RAG) (RAG faithfulness and answer relevance metrics): RAG metrics scoring how grounded an answer is in its context, how relevant it is to the question and how relevant the context is.
  • Ewaluatory kodowe (Code-based evaluators): A deterministic check in code when a failure can be described by a rule. The cheapest rung of evaluation — before an LLM judge.
    • Asercje deterministyczne: format, schemat, dopasowanie tekstu (Deterministic assertions: format, schema and string checks): Code graders that check outputs with rules: exact or regex match, schema validation, length and keyword presence.
    • Ocena przez uruchomienie testów (poprawność funkcjonalna) (Execution-based grading with unit tests): Grading generated code by executing it in a sandbox against unit or integration tests.
  • Interpretowalność (Interpretability and explainability): How to understand why a model made a given decision: from models that are transparent by design, through explanations added from outside (SHAP, saliency maps), to looking inside the network.
    • Interpretowalność mechanistyczna (Mechanistic interpretability): Studying a model's internal components and computations and testing how they contribute to its behaviour.
      • Obwody i grafy atrybucji (circuit tracing) (Circuits and attribution graphs): Building an approximate graph of features and connections that may mediate the computation behind a model answer. The graph requires intervention-based validation and does not guarantee a complete, faithful reconstruction of the mechanism.
      • Podmiana aktywacji (activation patching) (Activation patching): An intervention experiment: activations of a chosen component are replaced with activations from another input and the output change is measured. This provides causal evidence whose interpretation depends on the experimental design and metric.
      • Rzadkie autoenkodery (SAE) i słowniki cech (Sparse autoencoders and dictionary learning): Decomposing model activations into many sparsely active features, some of which can be associated with understandable concepts. Feature interpretability and practical downstream usefulness are not guaranteed.
    • Mapy istotności dla sieci neuronowych (Saliency maps and gradient attribution): Visualisations assigning importance to pixels or tokens for a network output using gradients or activations. High importance is not by itself evidence of causal influence.
    • Modele interpretowalne z natury (Inherently interpretable models): Models simple enough that their decision can be read directly — from weights, rules or a tree path — without extra explanation tools.
    • Ogólne zachowanie modelu: ważność cech i wykresy zależności (Global model explanations: feature importance and dependence plots): A description of model behaviour across a dataset: which features matter most to predictions and how model output varies as a feature changes. Such plots do not by themselves establish causality.
    • Sondy klasyfikacyjne (probing) (Probing classifiers): A classifier trained on a model's internal representations to test whether a property can be decoded from them. Strong probe performance alone does not prove that the model uses that information in its behaviour.
    • Sterowanie aktywacjami (wektory sterujące) (Activation steering and representation engineering): Finding directions or representations in activations associated with a concept and amplifying or suppressing them at run time to study or change model behaviour without retraining.
    • Wyjaśnianie pojedynczej decyzji (SHAP, LIME) (Local feature attribution (SHAP, LIME)): Methods that explain an individual prediction by assigning contributions to input features. They may treat the model as a black box or exploit its structure.
    • Wyjaśnienia „co by było, gdyby” (kontrfaktyczne) (Counterfactual explanations): The smallest change to the input that would flip the model's decision — e.g. what would have to change in an application for the loan to be approved.
  • Metryki i benchmarki (Metrics and benchmarks): Choosing a measure that fits the task and comparing models reliably: classic metrics, standardised benchmarks, leaderboards and statistical uncertainty in results.
    • Benchmarki modeli językowych (LLM benchmarks): Standardised task sets used to compare language models. They are grouped by the skill measured, such as knowledge and reasoning, coding or agent work.
      • Benchmarki agentowe (Agent benchmarks): Environments where a model carries out multi-step tasks with tools — in a browser, an operating system or a conversation with a simulated customer — and is scored on whether the goal was achieved.
      • Benchmarki długiego kontekstu (igła w stogu siana i dalej) (Long-context benchmarks (needle-in-a-haystack and beyond)): Synthetic and realistic tests of effective context length: retrieving planted facts, multi-hop tracing and aggregation.
      • Benchmarki programistyczne (Coding benchmarks): Tasks in which a model writes or fixes code and tests check the result — from single functions to real issues in repositories.
      • Benchmarki wiedzy i rozumowania (Knowledge and reasoning benchmarks): Closed-ended questions across many fields, from school level to expert level, graded automatically. Older sets saturate quickly, so ever harder ones are created.
      • Przeciek danych i nasycenie benchmarków (Benchmark contamination and saturation): Why a benchmark score can mislead: test questions may have leaked into training data, and popular tests stop separating top models once all score near 100%.
    • Metryki klasyfikacji i regresji (Classification and regression metrics): Classic measures for models that predict a class or a number. The choice depends on the cost of errors and whether classes are balanced.
    • Niepewność statystyczna wyników ewaluacji (Statistical uncertainty in evaluations): An evaluation is an experiment on a sample of tasks, so its score is uncertain. Without estimating that uncertainty, a small gap between models may be indistinguishable from sampling variation.
    • Porównanie tekstu z wzorcem (BLEU, ROUGE, BERTScore) (Reference-based text generation metrics): Automatic scoring of generated text by comparing it to a reference answer — from counting shared words to meaning similarity in embeddings.
    • Rankingi z głosami ludzi (areny) (Human-preference leaderboards (arenas)): People anonymously compare two model answers and choose the better one; many such comparisons produce a ranking. The result nevertheless depends on vote collection, models' access to the arena and comparison sampling.
  • Monitoring produkcyjny (Monitoring and observability): Watching a system after deployment: traces, cost, regressions.
    • Ewaluacja online na ruchu produkcyjnym (Online evaluation on production traffic): Continuous grading of sampled production traffic with code graders and LLM judges, complementing offline evals.
    • Wykrywanie dryfu danych i pojęć (Data and concept drift detection): Detecting that the production-data distribution or the relationship between inputs and correct outputs has changed relative to training data, which may require diagnosis or model adaptation.
    • Śledzenie wywołań modeli i agentów (tracing) (LLM and agent tracing): Instrumenting LLM applications to record model, tool and agent calls as a tree of spans with attributes.
  • Ocena ludzi i modeli (Human evaluation and model judges): Judging answers, with quality control on the judging itself.
    • Model językowy jako sędzia (LLM-as-a-judge) (LLM-as-a-judge): Using a language model as an automatic evaluator in single-answer, pairwise or reference-guided mode.
      • Uprzedzenia sędziów LLM: kolejność, długość, własny styl (LLM judge biases and mitigation): Systematic errors of LLM judges (position, verbosity, self-preference) and techniques to reduce them.
    • Ocena według rubryki i listy kontrolnej (Rubric-based and checklist grading): Grading answers against explicit criteria with level descriptions or binary checklist items, done by people or an LLM judge.
    • Walidacja ewaluatorów (Evaluator validation against human labels): Testing an LLM judge or classifier against human labels, on a set that played no part in tuning it.
  • Odporność i niepewność (Robustness and calibration): Behaviour on unusual data, and how trustworthy the confidence is.
    • Kalibracja pewności modelu (Confidence calibration): Whether a predicted probability or other confidence score matches how often the model is right — e.g. whether 80%-confidence predictions are correct about 80% of the time — and how to correct a mismatch.
    • Predykcja konformalna (Conformal prediction): Methods that equip any model with prediction sets or intervals having controlled marginal coverage under exchangeable data, without parametric distribution assumptions.
    • Wykrywanie halucynacji (Hallucination detection): Checking whether a model is inventing facts: against the provided source, against outside knowledge, or by how much its answers vary when the question is repeated.
    • Wykrywanie przykładów spoza rozkładu (OOD) (Out-of-distribution detection): Recognising inputs that do not fit the training-data distribution, so the system can raise an alert, abstain or hand the decision to a person.
  • Zbiory testowe i regresje (Evaluation datasets and regression tests): Repeatable checks on behaviour after changes.
    • Ewaluacja agentów zgodna z przepływem zadań (Task-flow aligned agent evaluation): Method turns issue-derived tasks into multi-turn interactions matching measured user task flows.
    • Frameworki do uruchamiania ewaluacji (Evaluation harnesses and frameworks): Tools that run test sets against a model or agent repeatably — with the same prompts, settings, sandbox and logs — so results can be compared and reproduced.
    • Syntetyczne przypadki testowe z wymiarów zapytań (Synthetic test case generation for evals): Building eval sets from query variation dimensions, value tuples and a separate conversion step into natural language.
• Fundamenty i rozumowanie (Foundations and reasoning): The mathematical and algorithmic language of AI, including methods without neural networks.
  • Logika i reprezentacja wiedzy (Logic and knowledge representation): How to write down facts, rules and dependencies.
    • AI neurosymboliczna (Neuro-symbolic AI): Combining neural networks (learning from data) with symbolic systems (logic, rules, solvers), e.g. an LLM translating a problem into a formal language for a solver.
    • Automatyczne dowodzenie twierdzeń (Lean) (Automated and formal theorem proving (Lean)): Models generate formal proofs checked by the kernel of a system such as Lean. Verification establishes that an accepted proof is correct relative to the formalized statement, declared axioms, and trusted kernel; the faithfulness of autoformalization remains a separate issue.
    • Grafy wiedzy (Knowledge graphs): Concepts and relations written as a graph, retrieval included.
    • Logika zdań i logika pierwszego rzędu (Propositional and first-order logic): Formal notation for facts and rules, and automatic derivation of conclusions from them.
    • Ontologie i sieć semantyczna (RDF, OWL) (Ontologies and the Semantic Web (RDF, OWL)): Standards for describing concepts, classes and relations so machines can check consistency and infer implicit facts.
    • Programowanie logiczne (Logic programming): Declarative paradigm: a program is a set of logical clauses and computation is inference (resolution, fixpoint or stable models).
    • Rozumowanie zdroworozsądkowe (Commonsense reasoning): AI subfield on representing and using everyday knowledge of the physical, social and temporal world.
    • Solwery SAT i SMT (SAT and SMT solvers): Engines deciding satisfiability of propositional formulas (SAT) and of formulas with background theories such as arithmetic or arrays (SMT).
  • Matematyka i optymalizacja (Mathematics and optimization): How to describe and optimise models.
    • Algebra liniowa: wektory i macierze (Linear algebra: vectors and matrices): The language for data, weights and embeddings: vectors, matrices, products and matrix decompositions.
      • Rozkłady macierzy: wartości własne i SVD (Matrix decompositions: eigendecomposition and SVD): Matrix factorisations (eigen, SVD, QR, Cholesky) revealing structure, rank and dominant directions of data.
    • Rachunek różniczkowy i gradienty (Calculus and gradients): Derivatives, gradients and the chain rule — the basis for computing how to change weights to reduce error (backpropagation).
    • Spadek gradientu i optymalizacja ciągła (Gradient descent and continuous optimisation): Mathematical foundations of minimising a loss: gradient descent, convexity, constraints. Practical training tuning lives under “Training optimization”.
      • Optymalizacja wypukła i dualność (Convex optimisation and duality): Minimising a convex function over a convex set, with Lagrangian duality theory and KKT optimality conditions.
    • Teoria informacji: entropia i dywergencja KL (Information theory: entropy, cross-entropy and KL divergence): Measures of information and of difference between distributions; cross-entropy is the standard loss of language models and classifiers.
  • Modele probabilistyczne (Probabilistic models): How to model dependencies and hidden variables.
    • Algorytm EM i modele mieszanin (Expectation-maximisation and mixture models): Iterative method for maximum-likelihood estimates in models with latent variables, such as Gaussian mixtures.
    • Pola losowe Markowa i grafy czynników (Markov random fields and factor graphs): Undirected graphical models and bipartite factor graphs, with inference by the sum-product algorithm (belief propagation).
    • Procesy gaussowskie (Gaussian processes): Non-parametric Bayesian model: a distribution over functions where any finite set of values is jointly Gaussian.
    • Programowanie probabilistyczne (Probabilistic programming): Languages and libraries where a generative model is written as code with sampling and conditioning statements, and inference runs automatically.
    • Sieci bayesowskie (Bayesian networks): A directed graph whose nodes are random variables and edges are dependencies; lets you compute probabilities from partial knowledge.
    • Ukryte modele Markowa i filtrowanie (Hidden Markov models and filtering): Models of hidden variables that change over time; tracking state from noisy observations.
    • Wnioskowanie przybliżone: próbkowanie i metody wariacyjne (Approximate inference: MCMC sampling and variational methods): When the exact posterior is intractable: approximate it with samples (MCMC) or by optimising a simpler distribution (VI).
    • Wnioskowanie przyczynowe (Causal inference): Telling causation from correlation: interventions, counterfactuals, confounders and discovering causal structure from data.
  • Przeszukiwanie i planowanie (Search and planning): How to find a route to a goal in a space of possibilities.
    • Gry i przeszukiwanie z przeciwnikiem (minimax) (Adversarial search and games (minimax, alpha-beta)): Choosing moves when an opponent plays against you: game trees, minimax, alpha-beta pruning, Monte Carlo tree search.
    • Planowanie automatyczne (PDDL) (Automated planning (PDDL)): Finding a sequence of actions that reaches a goal from a formal world description; today also as verifiable support for LLM agent planning.
    • Problemy z ograniczeniami (CSP) (Constraint satisfaction problems (CSP)): Assigning values to variables so all constraints hold — schedules, layouts, sudoku.
    • Przeszukiwanie heurystyczne (A*) (Heuristic search (A*)): A* orders nodes by the sum of path cost so far and heuristic cost to the goal. An admissible heuristic guarantees optimality in tree search; graph search must additionally handle newly found cheaper paths to previously visited states correctly.
    • Przeszukiwanie lokalne (wspinaczka, wyżarzanie) (Local search (hill climbing, simulated annealing)): Improving a candidate solution step by step by exploring its neighbours instead of the whole space: hill climbing, simulated annealing, tabu search.
    • Przeszukiwanie ślepe (BFS, DFS) (Uninformed search (BFS, DFS, uniform-cost)): Systematically exploring a state space without knowing where the goal is.
  • Statystyka i niepewność (Statistics and uncertainty): How to reason from incomplete data.
    • Estymacja parametrów (największa wiarygodność) (Parameter estimation (maximum likelihood, MAP)): Maximum likelihood estimation chooses parameters that maximize the likelihood of the observed data. In typical probabilistic classification and language models, minimizing the corresponding cross-entropy is equivalent to maximizing log-likelihood.
    • Rozkłady prawdopodobieństwa (Probability distributions): Random variables, distributions, expectation and variance; conditional probability and independence.
    • Testy statystyczne i przedziały ufności (Hypothesis tests and confidence intervals): Hypothesis tests and confidence intervals help assess observations relative to sampling variation and describe uncertainty. They do not by themselves determine whether an effect is true or practically important; they are used, among other things, in model comparisons, benchmarks, and A/B tests.
    • Twierdzenie Bayesa i wnioskowanie bayesowskie (Bayes' theorem and Bayesian inference): Updating beliefs after seeing data: prior, likelihood, posterior.
  • Teoria gier i decyzji (Game theory and decision theory): Theory of rational choice for a single decision-maker (utility) and many interacting players (equilibria, mechanism design, social choice).
    • Gry kooperacyjne i wartość Shapleya (Cooperative games and the Shapley value): Cooperative-game solution assigning each player their average marginal contribution over all orders of coalition formation.
    • Oczekiwana użyteczność i decyzje w niepewności (Expected utility and decisions under uncertainty): Normative theory of choice: a rational agent picks the action maximising the expected value of a utility function over outcomes.
    • Projektowanie mechanizmów i aukcje (Mechanism design and auctions): Reverse game theory: constructing rules (a mechanism) so players' strategic behaviour achieves a set goal, e.g. efficient allocation.
    • Równowaga Nasha i rozwiązywanie gier (Nash equilibrium and solving games): Game solution concept: a strategy profile in which each player's strategy is a best response to the others'.
    • Wybór społeczny: łączenie preferencji wielu osób (Computational social choice): Formal analysis of rules aggregating individual preferences into a group decision, with axioms, impossibility theorems and computational complexity.
  • Teoria uczenia statystycznego (Statistical learning theory): Theory of generalisation: error on unseen data, hypothesis-class complexity and the sample size needed to learn.
    • Brak darmowego obiadu i założenia wstępne modelu (No free lunch theorem and inductive bias): Result stating that, averaged over all possible problems, no learning algorithm beats another; success requires inductive bias.
    • Podwójne zejście (Double descent): Non-monotonic test-risk curve: classical U shape, a peak at the interpolation threshold and a second descent in the overparameterised regime.
    • Uczenie PAC i wymiar VC (PAC learning and VC dimension): Formal learnability model: with probability at least 1−δ the hypothesis error is ≤ ε, with sample size depending on the class's VC dimension.
• Inferencja i infrastruktura (Inference and ML systems): Running, accelerating and maintaining models.
  • AI lokalnie i na urządzeniach (Edge and on-device AI): Running on limited hardware and without a permanent cloud.
    • Eksport modeli i środowiska uruchomieniowe na urządzeniach (Model export and on-device runtimes): Moving a trained model from a training framework into a lightweight runtime on a phone, browser or embedded device: exporting or converting it to a supported format and delegating compute to the CPU, GPU or NPU.
    • Modele językowe na telefonie i laptopie (On-device language models): Language models optimised to run locally on a phone or laptop, often using an NPU, quantisation and other memory-saving techniques. On-device processing can keep data local and support offline use; hybrid architectures may route selected requests to a cloud model.
    • TinyML — modele na mikrokontrolerach (TinyML on microcontrollers): Running neural networks on microcontrollers with kilobytes of memory and milliwatts of power: co-designing tiny architectures with a lean inference engine, and quantising to 8 bits or fewer.
  • Akceleratory (AI hardware): Hardware for inference and training.
    • Model roofline: obliczenia czy przepustowość pamięci (Roofline model and arithmetic intensity): A simple performance model: attainable throughput is bounded by the minimum of peak compute and memory bandwidth multiplied by arithmetic intensity (operations per byte). Within the model's assumptions, it helps determine whether an operation is compute-bound or memory-bandwidth-bound.
    • Pamięć HBM i ściana pamięci (High-bandwidth memory and the memory wall): Accelerator compute grows faster than memory bandwidth, so moving weights and KV cache from HBM can become a bottleneck. This is especially important during LLM token generation, particularly at small batch sizes.
    • Połączenia między układami i sieć klastra (Accelerator interconnects and cluster networking): Ways accelerators exchange data: fast links within a server or pod (scale-up) and networking between servers (scale-out), using topologies such as torus or fat-tree. Their bandwidth and latency strongly affect how training and inference scale across many chips.
    • Wyspecjalizowane akceleratory AI (TPU, NPU) (Specialised AI accelerators (ASICs and NPUs)): Specialised chips or compute blocks optimised for neural-network operations, often using multiply-accumulate arrays and low-precision arithmetic. They include datacenter accelerators such as TPUs and NPUs integrated into phones and laptops.
  • GPU i obliczenia rozproszone (GPU and distributed computing): Using accelerators and many devices.
    • Języki i kompilatory kerneli (Kernel languages and compilers): Writing kernels one level above raw CUDA: tiles, templates, compilers.
    • Kernele mnożenia macierzy (GEMM) (Matrix multiplication kernels (GEMM)): The heart of AI compute: squeezing near-peak hardware throughput out of matrix multiplication.
    • Komunikacja między GPU (Multi-GPU communication): Moving data between many GPUs: collective operations and topology.
    • Profilowanie i optymalizacja kerneli (Kernel profiling and optimisation): Measuring where time goes, and ways to win it back.
    • Programowanie GPU (CUDA) (GPU programming (CUDA)): The GPU execution model: threads, blocks, the memory hierarchy.
    • Rdzenie tensorowe i niska precyzja (Tensor cores and low precision): Matrix-multiply units and the number formats that feed them.
    • Strategie zrównoleglania treningu (Training parallelism strategies): How to split a model and its data across devices.
  • Kwantyzacja i kompresja (Quantization and compression): Cutting memory and compute cost.
    • Kwantyzacja samych wag (Weight-only quantization): Storing weights in low precision with layer-by-layer calibration to keep quality at 4 bits and below.
    • Kwantyzacja wag i aktywacji (Weight and activation quantization): Lowering precision for activations too. Requires taming outlier values that break the scale.
    • Przycinanie wag i rzadkość (pruning) (Pruning and sparsity): Setting selected weights to zero to obtain a sparse model; pruning may be unstructured or follow a hardware-friendly pattern such as 2:4. The quality-sparsity trade-off depends on the model, calibration data and method; SparseGPT and Wanda are examples of one-shot LLM pruning without retraining.
  • MLOps i LLMOps (MLOps and LLMOps): Keeping deployments repeatable and observable.
    • Automatyczne retrenowanie i CI/CD dla ML (Continuous training and CI/CD for ML): Automating an ML pipeline: retraining and redeploying a model after a defined trigger such as a schedule, new data, quality degradation or detected drift, while testing and deploying the pipeline itself like code.
    • Bramka LLM: jedno API do wielu modeli (LLM gateway (AI gateway)): An intermediate layer in front of many providers and self-hosted models: one unified API, central keys and access control, rate limits and budgets, fallbacks, load balancing and cost logging.
    • Magazyn cech (feature store) (Feature store): A shared feature platform for training and production: an offline store with feature history for building training sets and an online store serving current features at low latency. Shared definitions and point-in-time joins help limit training-serving skew and leakage from the future.
    • Rejestr i wersjonowanie modeli (Model registry and versioning): A central registry of model versions with metadata and lineage to experiment runs. Mutable aliases such as champion or production can decouple production-version selection from deployment code and make promotion or rollback easier.
    • Wersjonowanie i rejestr promptów (Prompt management and versioning): Storing prompts as immutable, numbered versions with metadata and a production alias. Linking versions to traces and evaluations supports comparison; when an application loads prompts dynamically through an alias, rollback is possible without redeploying code.
    • Śledzenie eksperymentów (parametry, metryki, artefakty) (Experiment tracking): Logging training runs — hyperparameters, metrics, artefacts, and code and data versions — to support result comparison, lineage and reconstruction of experiment conditions.
  • Serwowanie i dekodowanie (Serving and decoding): Handling requests and generating responses.
    • Ciągłe batchowanie (Continuous batching): Adding requests to a running batch instead of waiting for it to finish.
    • Dekodowanie spekulatywne (Speculative decoding): A cheap model guesses several tokens ahead; the big one only verifies them.
    • Metryki wydajności inferencji LLM (TTFT, TPOT, przepustowość) (Inference performance metrics (TTFT, TPOT, throughput, goodput)): Core measures of LLM serving speed: time to first token (TTFT), average time for subsequent output tokens (TPOT/ITL), throughput in tokens or requests per second, and goodput — requests served within defined SLOs.
    • Rozdzielenie fazy prefill i decode (Disaggregated prefill and decode): Running prompt processing (prefill, compute-bound) and token generation (decode, memory-bound) on separate GPU pools, transferring the KV cache between them, so TTFT and per-token latency can be tuned independently.
    • Serwowanie wielu adapterów LoRA na jednym modelu (Multi-LoRA serving): One base model in GPU memory serves requests for many fine-tuned variants at once: a LoRA adapter is selected per request, and adapters are loaded and swapped on the fly.
    • Strategie dekodowania (Decoding strategies): How the next token is picked from the model's distribution.
    • Zarządzanie KV cache (KV cache management): The attention cache during generation: paging, sharing, compression.
  • Uruchamianie własnych modeli (Self-hosted model inference): Serving a model with your own weights from a local runtime or server.
    • Autoskalowanie i zimny start wdrożeń (Autoscaling model deployments and cold starts): Matching the number of model replicas to bursty traffic, including scale-to-zero. For GPU deployments, one major challenge is cold start: provisioning a machine, pulling the image and loading model weights.
    • Lokalne uruchamianie LLM na własnym komputerze (Running LLMs locally on consumer hardware): Running open models on a laptop or desktop computer. One common path uses quantised GGUF files and selects a quantisation level that fits available memory; a local API server is an optional way to expose the model to applications.
    • Silniki serwowania modeli językowych (LLM) (LLM inference engines and servers): Server software for serving LLMs efficiently to many users. Depending on the implementation, it may combine continuous batching, paged KV cache and shared-prefix reuse, and often exposes APIs compatible with popular interfaces plus metrics.
    • Szacowanie pamięci GPU (wagi i KV cache) (Estimating GPU memory for deployment): Working out how much GPU memory a deployment needs: weights (parameters × bytes per parameter at a given precision) plus a KV cache that grows with context length and concurrent requests, plus headroom for overhead.
• Język, obraz, audio (Language, vision and multimodality): Areas by signal type and task. They connect back to training, architectures and applications.
  • Dane przestrzenne i 3D (Spatial and 3D perception): Representations of the world and objects in space.
    • Chmury punktów i dane z LiDAR-u (Point clouds and LiDAR data): Machine-learning methods for sets of 3D points from LiDAR or scanners. PointNet processes points in a permutation-invariant way, while other approaches group them into voxels or vertical pillars; these representations support classification, segmentation and 3D object detection.
    • Gaussian splatting: szybkie sceny 3D ze zdjęć (3D Gaussian splatting): A scene representation built from anisotropic 3D Gaussians optimised from images and rendered by splatting. It enables high-quality real-time novel-view synthesis and provides an alternative to radiance-field representations.
    • Generowanie obiektów 3D z tekstu lub zdjęcia (Text-to-3D and image-to-3D generation): Creating 3D objects or scenes — such as meshes, Gaussians or radiance fields — from text or images. Methods include distilling guidance from 2D generators, generating consistent multiple views, and models trained directly on 3D asset datasets; outputs often require further editing.
    • Jednoczesna lokalizacja i budowa mapy (SLAM) (Simultaneous localisation and mapping (SLAM)): Simultaneously estimating the pose of a camera or robot while building or updating a map of an unknown environment. Classical methods combine observations with geometric optimisation, while newer systems also use learned modules, recurrent networks and scene representations based on radiance fields or Gaussians.
    • Neuronowe pola promieniowania (NeRF) (Neural radiance fields (NeRF)): Storing a 3D scene in a neural network that returns colour and density for a point in space and a viewing direction; trained from photos with known camera poses, it renders new, unphotographed views.
    • Szacowanie głębi z jednego zdjęcia (Monocular depth estimation): Predicting a depth map — each pixel's distance from the camera — from a single image. The output may represent only relative depth ordering or metric distance; training on large, diverse datasets has improved generalisation to new scenes.
  • Generowanie obrazu i wideo (Image and video generation): Synthesising and editing visual content.
    • Edycja obrazu poleceniem i podmiana fragmentów (inpainting) (Instruction-based image editing and inpainting): Changing an existing image according to a text instruction, regenerating a masked region (inpainting), or extending the image beyond its original frame (outpainting). Contemporary models can combine generation and editing, but preserving object identity across repeated edits still depends on the model and task.
    • Generowanie obrazów z opisu tekstowego (Text-to-image generation): Creating an image from a text description. Many contemporary systems use latent diffusion or transformers trained with flow-based objectives. Evaluation includes prompt adherence and image quality; some tools expose settings such as guidance scale, negative prompts or a random seed.
    • Generowanie wideo z tekstu i obrazu (Text-to-video and image-to-video generation): Creating clips from text descriptions or animating an image while trying to preserve motion and appearance across frames. Many contemporary systems use latent diffusion models and transformers that process space and time; some also generate synchronised audio.
    • Personalizacja: nauczenie modelu konkretnej osoby, rzeczy lub stylu (Personalisation: teaching a model a specific subject or style): Teaching a pretrained image model a new person, object or style from a few photos so it can be invoked with a special word in prompts. Methods: full fine-tuning (DreamBooth), a new word embedding (Textual Inversion), lightweight LoRA weights.
    • Sterowanie kompozycją: poza, kontury, głębia, obraz wzorcowy (Controllable generation: pose, edges, depth, reference image): Guiding generation with signals beyond text: edge maps, depth maps, pose skeletons, segmentation or a reference image. A common approach uses additional adapters or conditioning networks attached to a frozen base model; some conditions can be combined.
  • Język naturalny (Natural language processing): Understanding and producing language.
    • Automatyczna poprawa gramatyki i pisowni (Grammatical error correction): Detecting and correcting grammatical, spelling and word-choice errors in text with minimal changes to content.
    • Klasyfikacja tekstu i analiza sentymentu (Text classification and sentiment analysis): Assigning one or more labels to a text: sentiment, topic, spam, intent. Done today with a fine-tuned encoder or without task training — NLI-based zero-shot classification or prompting an LLM.
      • Klasyfikacja tekstu bez przykładów (zero-shot) (Zero-shot text classification): Assigning texts labels unseen in training, usually by recasting classes as NLI hypotheses or as an LLM prompt.
      • Opinie o konkretnych cechach (sentyment aspektowy) (Aspect-based sentiment analysis): Extracting aspect-level opinions from text: aspect term, aspect category, opinion term and polarity.
      • Wnioskowanie z tekstu: czy zdanie wynika z innego (NLI) (Natural language inference (textual entailment)): Classifying a premise–hypothesis pair as entailment, contradiction or neutral.
    • Odpowiadanie na pytania (QA) (Question answering (QA)): Answering natural-language questions. Extractive QA selects an answer span from a supplied text; open-domain QA first retrieves relevant sources and then derives an answer. This retriever-reader pattern is closely related to later RAG systems.
      • Odpowiadanie na pytania o dane w tabelach (Table question answering): Answering natural-language questions from a table: selecting cells, aggregating, or generating a query such as SQL.
      • Odpowiadanie na pytania z dużego zbioru tekstów (open-domain QA) (Open-domain question answering (retriever–reader)): Answering questions with no given context by retrieving passages from a large corpus and reading the answer from them.
      • Wskazywanie odpowiedzi w podanym tekście (extractive QA) (Extractive question answering (reading comprehension)): Answering a question by selecting an answer span in a given context, or stating that the context has no answer.
    • Streszczanie tekstu (Text summarisation): Shortening a text while preserving its key information. Extractive summarisation selects source passages, while abstractive summarisation creates new wording; modern language models are widely used for abstractive summaries, where quality evaluation and faithfulness to the source remain important challenges.
      • Streszczanie przez wybór najważniejszych zdań (Extractive summarisation): Building a summary from selected source sentences or spans, without generating new text.
      • Streszczanie własnymi słowami modelu (Abstractive summarisation): Generating a summary as new text with a sequence-to-sequence model or LLM, allowing paraphrase and information fusion.
    • Tokenizacja tekstu (BPE, WordPiece, SentencePiece) (Text tokenisation (BPE, WordPiece, SentencePiece)): Turning text into a sequence of tokens — sub-word pieces from a fixed vocabulary — before the model sees it. Sub-word methods dominate: BPE (also byte-level, as in GPT), WordPiece (BERT) and Unigram via SentencePiece (T5). Tokenisation affects cost, context length and how well a model handles non-English languages.
    • Tłumaczenie maszynowe (Machine translation): Automatic translation of text between languages. It evolved from statistical systems to neural encoder–decoders with attention and Transformers, then to massively multilingual models and prompted LLMs.
      • Ocena jakości tłumaczenia bez wzorca (quality estimation) (Machine translation quality estimation): Predicting machine-translation quality from the source and translation alone, without a reference translation.
      • Tłumaczenie mowy na inny język (Speech translation): Translating source speech into target-language text or speech, either cascaded (ASR→MT→TTS) or directly with one model.
    • Wielojęzyczność i języki o małych zasobach (Multilingual and low-resource language processing): Processing many languages at once and languages with little data. It relies on cross-lingual transfer in shared multilingual models; quality for low-resource languages still lags clearly behind English.
    • Wydobywanie informacji z tekstu (nazwy własne, relacje) (Information extraction (named entities and relations)): Turning free text into structured facts: named entities (NER — people, places, organisations, dates), relations and events. Classically done as token labelling; now also by LLMs emitting schema-constrained JSON or by compact zero-shot extractors.
      • Rozpoznawanie nazw własnych (NER) (Named entity recognition (NER)): Labelling text spans with entity types such as person, organisation, location or date; classically framed as token classification with a BIO scheme.
      • Rozwiązywanie koreferencji (kto to „on”?) (Coreference resolution): Clustering mentions (names, descriptions, pronouns) that refer to the same entity within a document.
      • Wydobywanie relacji między encjami (Relation extraction): Identifying typed relations between entity pairs in text and recording them as (subject, relation, object) triples.
      • Łączenie wzmianek z bazą wiedzy (entity linking) (Entity linking): Assigning a text mention the identifier of a knowledge-base entity, disambiguated by context.
  • Modele multimodalne (Multimodal models): Processing several kinds of data together.
    • Modele audio–język (rozumienie mowy, dźwięków i muzyki) (Audio-language models): Language models connected to audio representations that can jointly handle speech, environmental sounds and music: transcription, translation, questions about recordings, audio captioning or speaker-related analysis. The supported task range depends on each model's data and design.
    • Modele wizja–język (opisywanie obrazów i pytania o nie) (Vision-language models (VLMs)): Models that combine images or video with text to generate textual answers: they describe scenes, answer questions about visual content, read documents or identify objects. They vary in scale and design, from on-device models to large systems used for interface analysis and visual reasoning.
      • Automatyczne opisywanie zdjęć (image captioning) (Image captioning): Generating a natural-language description of an image with a model joining an image encoder and a text decoder.
      • Modele językowe rozumiejące wideo (Video-LLM) (Video-language models (Video-LLMs)): Models joining a video-frame encoder with an LLM for open-ended video understanding: description, questions, temporal localisation and reasoning.
      • Pytania i odpowiedzi o treść obrazu (VQA) (Visual question answering (VQA)): Answering a natural-language question about an image's content, open-ended or as multiple choice.
      • Rozumienie dokumentów, wykresów i zrzutów ekranu (Document, chart and screenshot understanding): Reading and interpreting images rich in text and spatial layout — scans, forms, invoices, charts, infographics and screenshots. Some models work without a separate OCR stage, answer questions or extract data directly from pixels; these capabilities are also used for interface analysis.
      • Wskazywanie na obrazie rzeczy opisanej słowami (visual grounding) (Visual grounding (referring expression comprehension)): Localising the image region (box or mask) matching a language expression, including expressions matching several objects or none.
      • Łączenie enkodera obrazu z modelem językowym (Connecting an image encoder to a language model): A common way to build a VLM: a pretrained image encoder produces visual features, while a projector, adapter or cross-attention layers expose them to a language model. The system is then tuned on image-instruction data; the modules trained and number of stages depend on the architecture.
    • Ujednolicone modele wielomodalne (wejście i wyjście) (Unified multimodal models with multimodal input and output): Models that handle several input modalities — such as text, images, audio and video — and at least several output forms, such as text or speech. They combine these signals in one system; architectures differ in tokenisation, modality fusion and output generation.
  • Mowa i audio (Speech and audio): Analysing and synthesising sound.
    • Agenci głosowi (Real-time voice agents): A system that conducts spoken conversations, understands utterances, makes decisions and completes tasks using models and tools.
      • Architektury agentów głosowych (Voice agent architectures): Ways to combine speech understanding, agent decisions and voice generation: modular, integrated or with delegated reasoning.
        • Głosowy front i osobny agent zadaniowy (Voice frontend with delegated reasoning): A voice model conducts the live conversation while delegating reasoning and tools to a separate text agent; the result returns to the spoken dialogue.
        • Modele mowa–mowa (Speech-to-speech models): A model takes speech and generates speech without separate STT and TTS services in the pipeline. Its internal representation may still include text or transcription; tool use and context control depend on the model.
        • Potok STT → agent tekstowy → TTS (Cascaded STT–text agent–TTS pipeline): Separate speech recognition, a text agent and speech synthesis. Stages can be swapped, inspected and connected to an existing workflow.
      • Ewaluacja agentów głosowych (Voice agent evaluation): Evaluating task success and conversational quality on real audio: tools, turns, interruptions, audible latency and robustness across speaking conditions.
        • Ocena tur, przerwań i słyszalnego czasu reakcji (Turn-taking, interruption and audible-latency evaluation): Tests measure premature takeovers, unwanted silence, overlap, post-interruption behavior and time to a useful audible response.
        • Odporność na akcent, hałas i cudzą mowę (Robustness to accents, noise and third-party speech): The same tasks are tested with varied accents, noise and other speakers, tracking understanding, turns and task outcomes separately.
        • Symulacje wieloturowych rozmów audio (Multi-turn audio conversation simulation): A simulated caller speaks with the agent through a full scenario, including pauses, corrections and interruptions; both dialogue and task state are evaluated.
        • Weryfikacja wyniku zadania i narzędzi (Voice-agent task and tool outcome verification): Evaluation compares spoken user intent, tool calls, final system state and the agent's spoken confirmation.
      • Generowanie odpowiedzi w czasie rzeczywistym (Real-time spoken response generation): Preparing agent speech so a useful response begins promptly while the model remains responsive to further user speech.
        • Strumieniowe przekazywanie odpowiedzi do mowy (Streaming model output into speech): Model output is passed to speech generation before the whole reply is complete; chunk boundaries affect latency and naturalness.
        • Wyprzedzające generowanie odpowiedzi (Preemptive response generation): The agent starts model work before the user's turn is confirmed complete; continued user speech may require discarding or revising the response.
      • Prowadzenie tur rozmowy (Turn-taking in voice agents): Decisions about when to listen, start responding and allow simultaneous listening and speaking. Silence does not necessarily mean a thought is complete.
        • Detekcja aktywności głosowej (VAD) (Voice activity detection): Detecting speech segments in an audio stream. VAD can trigger downstream stages but does not itself determine whether the user has finished a turn.
        • Dialog pełnodupleksowy (Full-duplex spoken dialogue): A system can receive speech and generate a response at the same time, handling overlap and listener backchannels. This can be achieved with an integrated model or by synchronizing modules.
        • Kontekstowe wykrywanie tury (Context-aware turn detection): Estimating whether the user has completed a thought from semantic and/or acoustic cues. Implementations may use audio, a transcript or both.
      • Przerwania i powrót do rozmowy (Interruption handling and recovery): Determining whether the user is interrupting, stopping the response and resuming the dialogue and task in line with what was already heard.
        • Rozróżnianie przerwania od potakiwania i tła (Distinguishing interruptions from backchannels and background speech): After detecting audio, deciding whether the user wants to take the turn, is backchanneling or coughing, or another person spoke in the background.
        • Usuwanie niesłyszanej części odpowiedzi z kontekstu (Aligning conversation context with played speech): After an interruption, remove generated speech the user did not hear from conversation history. The boundary depends on playback tracking and the API.
        • Wznowienie zadania po przerwaniu (Task recovery after a voice interruption): After the user interrupts, the agent handles the new request and returns to the right task step without repeating speech already heard.
      • Rozumienie wypowiedzi w dialogu (Understanding spoken input in voice agents): Interpreting who is speaking and what they said, including uncertainty and exact data capture. ASR itself is a separate topic in the map.
        • Dopytywanie przy niepewnej interpretacji mowy (Clarification under spoken-input uncertainty): When speech admits multiple interpretations, the agent asks a short clarifying question rather than guessing or taking action.
        • Przechwytywanie i potwierdzanie dokładnych danych (Exact entity capture from speech): The agent captures spoken names, numbers and codes, clarifies ambiguous characters and reads back exact values for confirmation before using a tool.
        • Rozpoznawanie właściwego rozmówcy (Target-speaker attribution in voice agents): Distinguishing the user's speech from other speakers so background voices do not become commands or interruptions.
      • Zadania i narzędzia w rozmowie (Task and tool orchestration in spoken dialogue): Using tools and completing multi-step tasks while keeping the conversation coherent. Calls may block or run in the background, depending on the API and agent design.
        • Komunikowanie pracy narzędzia (Spoken progress cues during tool use): The agent indicates it is working or waiting for a result while keeping its speech aligned with the actual operation state.
        • Narzędzia działające w tle rozmowy (Non-blocking tools in spoken dialogue): The agent can keep conversing while a long-running task executes, then incorporate its result. Semantics and support vary across APIs.
        • Potwierdzanie działań wykonywanych głosem (Spoken confirmation before consequential actions): Before an action that changes data, the agent speaks the intended action and consequence, waits for explicit confirmation and reports success only after the tool succeeds.
        • Scenariusze i stan rozmowy (Stateful spoken workflows): The model conducts a multi-step conversation; changing the step affects instructions, available tools and transferred context.
    • Automatyczny zapis nut z nagrania (Automatic music transcription): This identifies the pitch and duration of notes heard in a recording.
    • Diaryzacja: kto kiedy mówi (Speaker diarization): This labels which parts of a conversation are spoken by the same person.
    • Dopasowanie tekstu do czasu nagrania (Forced alignment of speech and text): This determines when words from a supplied text occur in a recording.
    • Generowanie muzyki i efektów dźwiękowych (Music and sound generation): Models that create music, sound effects or ambience from a text prompt (sometimes with a melody or lyrics), usually as language models over discrete audio tokens or as latent diffusion.
    • Oczyszczanie i poprawa jakości mowy (Speech enhancement): This reduces interference to make a recorded voice clearer.
    • Rozdzielanie wokalu i instrumentów (Music source separation): This splits a finished music recording into separate sound parts.
    • Rozpoznawanie dźwięków i klasyfikacja nagrań (Audio and sound event classification): Assigning one or more labels that describe a recording, such as sound type, music genre, spoken command or language. Models often analyse a spectrogram or a learned audio representation; shared audio-text spaces also enable zero-shot classification.
    • Rozpoznawanie mowy / ASR (Automatic speech recognition): Turning speech into text.
    • Synteza mowy / TTS (Text-to-speech): Producing speech from text.
  • Widzenie komputerowe (Computer vision): Recognising content and structure in images and video: what is in the picture, where it is, what shape it has and how it moves. Image generation, 3D and vision–language models have their own sibling nodes.
    • Detekcja obiektów (Object detection): Finding objects in an image: a bounding box plus a class for each object. Usually evaluated with IoU and mAP.
      • Detekcja obiektów opisanych tekstem (Open-vocabulary object detection): Detecting objects queried by names or natural-language descriptions rather than one fixed closed label set; includes zero-shot generalisation to new categories.
      • Detekcja z transformerem (DETR) (Transformer-based object detection (DETR)): Detection as direct set prediction by a transformer, without hand-designed anchors or NMS post-processing.
      • Detektory jednoetapowe (YOLO) (Single-stage object detectors): The model predicts boxes and classes without a separate region-proposal stage. The YOLO family shows that such detectors can run in real time and on edge devices.
    • Estymacja pozy i punktów kluczowych (Pose and keypoint estimation): Detecting keypoints (joints, hands, face) and body skeletons in 2D or 3D — the basis of motion analysis, sports and gesture control.
    • Klasyfikacja obrazów (Image classification): Assigning one or more labels to a whole image (e.g. “cat”, “defective product”). It is one of the foundational computer-vision tasks.
    • Modele bazowe dla obrazu (Vision backbones and foundation models): Pretrained vision models that extract image features reusable for classification, detection and segmentation.
      • Dostrajanie modeli wizyjnych do własnych danych (Fine-tuning vision models on custom data): Adapting a pretrained model to a custom dataset by fine-tuning the whole network, selected layers, or a new task head.
      • Samonadzorowane uczenie z obrazów (Self-supervised visual pretraining): Training a vision model on unlabelled images — e.g. by reconstructing masked patches or comparing different views of the same image.
      • Vision Transformer (ViT): A transformer applied to an image split into patches treated as tokens; a major family of vision architectures alongside convolutional networks.
      • Wspólna przestrzeń obraz–tekst (CLIP) (Contrastive image–text pretraining (CLIP)): Learning joint embeddings of images and captions from image–text pairs; enables zero-shot classification and image search by text description.
    • Powiększanie i naprawa zdjęć (super-rozdzielczość) (Image upscaling and restoration (super-resolution)): Recovering a clean, higher-resolution image from a degraded one (low resolution, noise, blur, compression). Approaches range from GAN models trained on synthetic degradations to diffusion-based restoration with a generative prior, which can add plausible but not necessarily true detail.
    • Rozpoznawanie tekstu na obrazach (OCR) (Optical character recognition (OCR)): Converting images of print, handwriting or scene text into machine text; usually line or word detection plus recognition.
    • Rozumienie wideo (Video understanding): Analysing frame sequences: adds time, motion and far more data than a single image.
      • Rozpoznawanie akcji w wideo (Action recognition): Recognising what happens in a clip (running, falling, passing a ball) from appearance and motion over time.
      • Śledzenie obiektów w wideo (Multi-object tracking): Keeping the identity of the same objects across frames, usually by linking detector outputs frame to frame (tracking-by-detection).
    • Segmentacja obrazu (Image segmentation): Dividing an image into regions at pixel level — masks instead of boxes.
      • Segmentacja instancji i panoptyczna (Instance and panoptic segmentation): A separate mask for each object (instances); the panoptic variant unifies objects and background in one consistent map.
      • Segmentacja na wskazanie (Segment Anything) (Promptable segmentation): A general model returns a mask for an object indicated by a click, a box or a short text phrase — no class-specific training; newer versions also work on video.
      • Segmentacja semantyczna (Semantic segmentation): Every pixel gets a class label (road, sky, person) without separating individual instances.
• Obliczenia ewolucyjne i inteligencja roju (Evolutionary computation and swarm intelligence): Population-based optimisation and search methods inspired by evolution and collective behaviour, working without gradients, only on candidate scores.
  • Budowa algorytmu ewolucyjnego: reprezentacja, operatory, parametry (Designing an evolutionary algorithm: representation, operators and parameters): Components of an evolutionary algorithm: representation, initialisation, parent and survivor selection, variation operators, constraint handling, parameter control and sound comparison.
    • Krzyżowanie: łączenie cech dwóch rodziców (Crossover (recombination) operators): Recombination operators building offspring from several parents' material, chosen to fit the representation: binary, real-valued or permutation.
    • Mutacja: drobne losowe zmiany rozwiązania (Mutation operators): Unary variation operators adding random perturbations to an individual; they provide exploration and reachability of the whole space.
    • Nisze i utrzymywanie różnorodności populacji (Niching and diversity preservation): Diversity-preserving techniques that let population-based metaheuristics locate multiple optima of a multimodal function in a single run.
    • Obsługa ograniczeń w algorytmach ewolucyjnych (kary, naprawa) (Constraint handling in evolutionary algorithms): Techniques for steering evolutionary search in constrained spaces: penalties, repair, feasibility rules, special operators and multi-objective formulations.
    • Reprezentacja osobnika (kodowanie binarne, liczbowe, permutacje, drzewa) (Solution representation in evolutionary algorithms): The genotype–phenotype mapping: choosing a data structure for candidates that defines the search space and admissible operators.
    • Rzetelne porównywanie algorytmów optymalizacji (benchmarki) (Benchmarking evolutionary and black-box optimisers): Methodology for comparing black-box optimisers: standardised function suites, evaluation budgets, time-to-target measures and statistical tests.
      • Krytyka „nowych” metaheurystyk opartych na metaforach (Critique of metaphor-based metaheuristics): A methodological strand showing many metaphor-based metaheuristics are renamed known techniques, and calling for description in standard optimisation terms.
    • Selekcja rodziców i przeżycia (turniej, ruletka, elitaryzm) (Parent and survivor selection (tournament, roulette, elitism)): Mechanisms for choosing parents and survivors that control selection pressure: proportional, rank-based and tournament, plus replacement strategies and elitism.
    • Strojenie i samoadaptacja parametrów algorytmu ewolucyjnego (Parameter tuning and control in evolutionary algorithms): Parameter tuning (static choice before a run) and parameter control (deterministic, adaptive or self-adaptive changes during a run).
  • Ewolucja z modelami językowymi (LLM-guided evolutionary search): Methods where large language models act as variation operators in an evolutionary algorithm while an evaluator automatically scores programs, prompts or agents.
    • Ewolucyjne odkrywanie algorytmów z LLM (LLM-driven program and algorithm evolution): Evolutionary search over programs in which a language model generates code variants and an evaluator scores them on a mathematical or engineering task.
    • Ewolucyjne ulepszanie poleceń dla modeli językowych (Evolutionary prompt optimisation): Optimising prompt text with an evolutionary algorithm where a language model performs mutation and crossover and selection uses validation scores.
    • Model językowy jako operator krzyżowania i mutacji (Language models as evolutionary variation operators): Using a language model as the variation operator in an evolutionary algorithm: mutation by editing or crossover by prompting several parents.
    • Samodoskonalące się agenty (ewolucja agentów) (Evolutionary self-improvement of agents): Evolutionary, archive-based search over agent designs where a foundation model edits agent code and benchmarks empirically validate changes.
  • Inteligencja roju (Swarm intelligence): Optimisation methods built on collective behaviour of many simple agents, mainly particle swarms and ant colony algorithms.
    • Algorytmy mrówkowe (ACO) (Ant colony optimisation (ACO)): A metaheuristic for combinatorial problems where agents probabilistically build solutions guided by pheromone and heuristics, with pheromone reinforced and evaporated.
    • Optymalizacja rojem cząstek (PSO) (Particle swarm optimisation (PSO)): A population metaheuristic in which particles update velocity from inertia, their personal best and their neighbourhood best.
    • Robotyka rojowa (Swarm robotics): Designing collective behaviour for large groups of simple autonomous robots with local sensing and communication and no central control.
    • Symulacja stada (boids) (Boids and flocking simulation): A distributed flocking model where each agent steers by three rules relative to neighbours within a local range.
    • Sztuczna kolonia pszczół (ABC) (Artificial bee colony algorithm (ABC)): A swarm metaheuristic with three bee roles — employed, onlooker and scout — for continuous optimisation.
  • Klasyczne algorytmy ewolucyjne (Classical evolutionary algorithms): The main evolutionary algorithm families: genetic algorithms, evolution strategies, genetic programming and evolutionary multi-objective optimisation.
    • Algorytmy estymacji rozkładu (EDA) (Estimation of distribution algorithms (EDA)): Probabilistic model-building metaheuristics: they fit a distribution to selected good solutions and sample the next population from it.
    • Algorytmy genetyczne (Genetic algorithms (GA)): A population algorithm encoding solutions as chromosomes and creating offspring through selection, crossover and mutation.
      • Model wyspowy (równoległe populacje z migracją) (Island model and parallel evolutionary algorithms): Coarse-grained parallel evolution: independent subpopulations exchanging individuals according to a migration topology and schedule.
    • Algorytmy memetyczne (ewolucja z lokalnym ulepszaniem) (Memetic algorithms): An evolutionary algorithm with one or more local search phases inside its generational cycle.
    • Ewolucja różnicowa (DE) (Differential evolution (DE)): A population method for continuous optimisation where a mutant is a base vector plus a scaled difference of two random members.
    • Ewolucja wspomagana modelem zastępczym (gdy ocena jest droga) (Surrogate-assisted evolutionary optimisation): Evolutionary algorithms replacing some expensive objective evaluations with approximations from regression or classification models learned during the search.
    • Ewolucyjna optymalizacja wielokryterialna (NSGA-II) (Evolutionary multi-objective optimisation (NSGA-II)): Evolutionary algorithms approximating the Pareto front of a multi-objective problem in one run, such as NSGA-II, NSGA-III and MOEA/D.
      • Optymalizacja przy wielu celach naraz (NSGA-III, MOEA/D) (Many-objective evolutionary optimisation (NSGA-III, MOEA/D)): Evolutionary algorithms for problems with typically four or more objectives, built on reference points or decomposition into scalar subproblems.
      • Wskaźnik hiperobjętości frontu Pareto (Hypervolume indicator for Pareto fronts): A quality indicator for a set of non-dominated solutions, measuring the objective-space volume dominated by that set relative to a reference point.
    • Koewolucja: populacje oceniane przez wzajemne starcia (Coevolutionary algorithms): Evolutionary algorithms in which at least part of the selection signal comes from interactions among individuals in one or more populations.
    • Programowanie genetyczne i regresja symboliczna (Genetic programming and symbolic regression): Evolution of expression trees or programs; its main use is symbolic regression, discovering readable formulas from data.
      • Ewolucja gramatyczna (programy budowane z reguł gramatyki) (Grammatical evolution): Genetic programming with a genotype–phenotype mapping through a context-free grammar in BNF notation, evolving programs in an arbitrary language.
      • Ewolucyjne ulepszanie i naprawa istniejącego kodu (Genetic improvement and automated program repair): Search, usually by genetic programming, over edits to existing code to repair bugs or improve performance.
      • Kartezjańskie programowanie genetyczne (programy jako siatka węzłów) (Cartesian genetic programming (CGP)): Genetic programming encoded as a directed graph on a grid of nodes, with an integer genotype and inactive genes enabling neutral drift.
    • Strategie ewolucyjne i CMA-ES (Evolution strategies and CMA-ES): Evolutionary continuous optimisers that sample candidates from a normal distribution and adapt its mean, step size and covariance matrix.
      • Naturalne strategie ewolucyjne (NES) (Natural evolution strategies (NES)): Evolution strategies that update search-distribution parameters by the natural gradient of expected fitness.
    • Uczące się systemy klasyfikatorów (ewolucja reguł „jeśli–to”) (Learning classifier systems (LCS)): Rule-based learning systems combining a population of condition–action classifiers, reward-based credit assignment and a genetic algorithm discovering new rules.
  • Neuroewolucja (Neuroevolution): Applying evolutionary algorithms to optimise weights, architectures, hyperparameters or learning rules of neural networks.
    • Ewolucja budowy sieci (NEAT) (NeuroEvolution of Augmenting Topologies (NEAT)): An algorithm evolving network weights and topology together, using historical gene markings, speciation and gradual complexification from minimal structure.
      • Ewolucja wzorów połączeń dużych sieci (HyperNEAT) (HyperNEAT and compositional pattern-producing networks): Neuroevolution with indirect encoding: NEAT evolves a CPPN that sets connection weights between neurons placed in geometric space.
    • Ewolucyjne projektowanie architektur sieci (NAS) (Evolutionary neural architecture search): Searching neural architecture space with an evolutionary algorithm whose fitness is the quality of the trained candidate.
    • Strategie ewolucyjne w trenowaniu sieci i LLM (Evolution strategies as an alternative to RL): Scalable evolution strategies estimating an improvement direction from scores of randomly perturbed parameter copies, used in RL and in fine-tuning large language models.
    • Trenowanie populacyjne (Population Based Training) (Population Based Training (PBT)): An asynchronous method jointly optimising weights and hyperparameters of a model population, discovering a hyperparameter schedule rather than one fixed setting.
  • Różnorodność i ewolucja otwarta (Quality diversity and open-endedness): Algorithms that illuminate a solution space with collections of diverse high-quality candidates, plus research on open-ended, never-ending generation of novelty.
    • MAP-Elites: mapa najlepszych wariantów (MAP-Elites): A quality-diversity algorithm keeping a grid of niches defined by behaviour features and the best solution in each niche.
    • Poszukiwanie nowości (novelty search) (Novelty search): An evolutionary algorithm replacing the objective with a behavioural novelty score relative to population and archive, effective on deceptive problems.
    • Red teaming jako wyszukiwanie quality-diversity (Quality-diversity red teaming): Generating diverse adversarial prompts as a quality-diversity problem: a MAP-Elites archive indexed by risk category and attack style.
    • Systemy otwarte (open-endedness) (Open-ended learning systems): Research on systems that endlessly generate novel, learnable artefacts or tasks along with their solutions.
• Produkt i wdrażanie AI (AI product and adoption): Turning model capability into something useful and into people's work.
  • Adopcja i edukacja (AI adoption and enablement): Bringing AI into a team's daily work.
    • Kompetencje AI w zespole (AI literacy) (AI literacy and upskilling): Training people who use or operate AI so they understand its capabilities, risks and their own role — tailored to the job and context; in the EU also an Article 4 AI Act requirement.
    • Pomiar wpływu AI na produktywność (Measuring AI productivity impact): Checking whether AI actually speeds up work with experiments or work data rather than self-reports — because users' impressions can diverge sharply from measurement.
    • Przeprojektowanie procesów pracy pod AI (Workflow redesign for AI): Changing how work is done — step order, roles and checkpoints — instead of bolting AI onto the old process; industry surveys name it the main difference between firms seeing impact and the rest.
    • Zasady korzystania z AI w firmie (shadow AI) (Organisational AI use policy and shadow AI): Defining which AI tools employees may use and with what data, and how to move from unofficial use of personal tools to approved solutions.
  • Dobór zastosowań (Use-case selection): Checking whether AI solves the right problem.
    • Automatyzacja czy wspomaganie człowieka (Automation vs augmentation): Deciding which tasks AI should do on its own and where it should only assist a person — weighing the cost of errors and what workers themselves want.
    • Budować czy kupić: API, model otwarty czy gotowe narzędzie (Build vs buy for AI capabilities): Choosing between an off-the-shelf product, a model via API, self-hosting an open-weights model and fine-tuning — by cost, control, data and the advantage it gives.
    • Ocena wartości i wykonalności zastosowania AI (AI use-case value and feasibility assessment): Checking whether AI solves a real problem better than a simpler alternative, what it is worth, and whether the data and model quality exist to make it work.
    • Prototypowanie funkcji AI (proof of concept) (AI feature prototyping): Quickly testing an idea with the strongest available model and a small set of examples before investing in optimisation, cheaper models or fine-tuning.
  • Eksperymenty produktowe (Product experimentation): Testing the real effect of a solution.
    • Metryki sukcesu produktu AI (AI product success metrics): Choosing measures that show whether an AI feature actually helps: one primary task-success criterion plus guardrail metrics that make sure nothing else breaks.
    • Testy A/B funkcji AI (A/B testing AI features): Randomly splitting users between versions of a model, prompt or interface and comparing them on real traffic — a complement to offline evaluation, not a replacement.
    • Wdrożenie w cieniu i stopniowe (canary) (Shadow and canary rollouts): Releasing model or prompt changes safely: first in shadow mode without showing outputs, then to a small share of users, with fast rollback if problems appear.
  • Koszt, jakość i opóźnienie (Cost, quality and latency trade-offs): Picking the trade-off that fits the use case.
    • Cache w aplikacjach LLM (Caching in LLM applications): Avoiding recomputing the same thing: storing repeated prompt parts or whole answers, which lowers cost and response time.
      • Cache promptu (wspólny początek zapytań) (Prompt (prefix) caching): Reusing an already processed fixed prompt prefix — system instructions, tool definitions, long documents — at a fraction of the price and with a shorter time to first token.
      • Cache semantyczny odpowiedzi (Semantic response caching): Returning a stored answer for a question similar in meaning to an earlier one — matched by embeddings rather than exact text; needs a similarity threshold and control of stale answers.
    • Kaskady modeli: najpierw tańszy model (LLM cascades): Sending a query to a cheaper model first and escalating to a more expensive one only when the answer looks uncertain or wrong — saving cost while keeping quality.
    • Przetwarzanie wsadowe (batch) zamiast na żywo (Batch (asynchronous) inference): Sending large bundles of requests that don't need an immediate answer in batch mode — usually much cheaper, at the cost of answers within hours instead of seconds.
    • Szacowanie kosztu na zadanie (tokeny) (Cost-per-task estimation): Working out what one run of a task costs: input and output tokens, number of model and tool calls, retries — and comparing it with the value the task delivers.
  • UX i współpraca z AI (Human–AI interaction): Designing control, feedback and trust.
    • Kalibracja zaufania: cytaty, pewność i wyjaśnienia (Trust calibration): Showing signals that let users judge when to trust an AI answer: sources, confidence levels, short explanations and labelling of generated content.
    • Kontrola użytkownika: sugestia, edycja, akceptacja (User control and oversight patterns): Patterns where AI proposes and the person edits and approves: a plan to accept, draft mode, variants to pick from, easy undo and a manual fallback path.
    • Odczuwalna szybkość: strumieniowanie i wskaźniki postępu (Perceived latency and progress feedback): Making waiting for AI tolerable: streaming text, showing agent work steps and allowing interruption — following the 0.1 s / 1 s / 10 s perception thresholds.
    • Projektowanie na wypadek błędów AI (Error handling and graceful failure): Assuming up front that the AI will be wrong and planning how users notice, correct and resume — in proportion to what is at stake.
    • Ustawianie oczekiwań: co AI potrafi, a czego nie (Setting expectations and mental models): Helping users understand what the system can do, how well, and where it fails — so they neither overtrust it nor abandon it after the first mistake.
    • Wzorce interfejsów AI (czat, copilot, akcja w tekście) (AI interface patterns): Recurring forms in which AI appears in a product — from a separate chat window, through a side assistant, to actions built into the text being edited — and when each works.
    • Zbieranie opinii użytkowników (jawnych i niejawnych) (Explicit and implicit user feedback): Designing ways for users to signal AI quality: direct ratings and comments, plus behavioural signals — acceptance, edits, regeneration, abandonment.
• Robotyka i embodied AI (Robotics and embodied AI): The field combining perception, planning, control and learning for machines acting in the physical world: arms, mobile robots and humanoids.
  • Dane i symulacja dla robotów (Robot data and simulation): Data sources for robot learning—teleoperation, physics simulation and human recordings—and methods for transferring between them.
    • Przenoszenie z symulacji do rzeczywistości (sim-to-real) (Sim-to-real transfer): Training in simulation and transferring to a real robot despite the differences (the “reality gap”), e.g. by randomising the simulation's appearance and physics.
      • Cyfrowy bliźniak sceny do trenowania robota (real-to-sim) (Real-to-sim-to-real with scene digital twins): Building a simulation from a real scene (reconstructing geometry, appearance and objects) and refining a policy there, e.g. with RL, before returning to hardware.
    • Teleoperacja i zbiory demonstracji robotów (Teleoperation and robot demonstration datasets): Collecting robot-learning data: a human drives the robot (e.g. leader–follower arms) and the recordings go into shared datasets spanning many robot types.
    • Testowanie polityk robota w symulacji i benchmarkach (Robot policy evaluation and benchmarks): Evaluating manipulation and navigation policies: simulated task benchmarks, simulations mirroring real setups, and hardware test protocols.
    • Uczenie robotów z nagrań ludzi (Learning robot skills from human videos): Pretraining visual representations on human video, often egocentric, to support later robot-manipulation learning.
  • Drony i roboty latające (Aerial robotics and autonomous drone flight): Autonomy for flying robots: state estimation from cameras and inertial sensors, flight planning and control, including policies learned in simulation.
    • Samodzielna nawigacja drona w ciasnych miejscach (Vision-based autonomous drone navigation): GPS-denied drone navigation from cameras and inertial sensors: localisation, mapping, obstacle-avoidance planning and learned policies.
    • Szybki lot drona sterowany siecią neuronową (wyścigi dronów) (Learned agile flight and autonomous drone racing): Flight policies learned by RL or imitation in simulation that issue low-level commands from onboard perception at speeds and accelerations near the drone's limits.
  • Floty i zespoły robotów (Multi-robot systems and fleet coordination): Coordinating many robots: task allocation, collision-free multi-agent path planning, formation control and swarms, and fleet management.
    • Planowanie tras wielu robotów bez kolizji (MAPF) (Multi-agent path finding (MAPF)): Computing collision-free paths for many agents from starts to goals: search-based (CBS, PBS, LNS), compilation-based (SAT, MIP) and learned methods.
    • Przydzielanie zadań robotom we flocie (Multi-robot task allocation (MRTA)): Assigning tasks to robots to minimise time, energy or fleet size: exact methods, heuristics, auctions and learned policies.
  • Kinematyka i sterowanie robotem (Robot kinematics and control): The classical core of robotics: describing motion geometry, planning collision-free trajectories and closed-loop actuator control.
    • Kinematyka prosta i odwrotna (Forward and inverse kinematics): The mapping between joint space and end-effector pose: direct one way, ambiguous and often iterative the other.
    • Planowanie ruchu ramienia bez kolizji (Collision-free motion planning for manipulators): Searching for a collision-free path in an arm's configuration space with sampling-based methods or trajectory optimisation.
    • Planowanie zadań i ruchów naraz (TAMP) (Task and motion planning (TAMP)): Planning that combines discrete action choice (symbolic planning) with continuous motion planning and parameter selection, e.g. grasps and placements.
    • Sterowanie siłą i podatnością (sterowanie impedancyjne) (Impedance, admittance and force control): Controlling the force–motion relationship: impedance (virtual spring-damper), admittance (force mapped to motion) and direct force control, including hybrid schemes.
    • Sterowanie ze sprzężeniem zwrotnym (PID, MPC) (Feedback control for robots (PID, MPC)): Closed-loop controllers—from classical PID through LQR to model predictive control—that track a trajectory despite disturbances.
  • Lokomocja i roboty humanoidalne (Legged locomotion and humanoid robots): Legged robots and humanoids: from model-based contact control to RL policies and human-motion imitation for loco-manipulation.
    • Lokomocja robotów uczona w symulacji (Simulation-trained legged locomotion): Walking policies for legged robots trained with RL across many parallel simulations and then transferred to physical hardware.
      • Chodzenie po trudnym terenie z patrzeniem pod nogi (Perceptive legged locomotion on rough terrain): Locomotion policies fusing proprioception with terrain perception (height map, depth images), trained by RL in simulation and robust to unreliable external sensors.
    • Naśladowanie ruchu człowieka przez robota (Motion imitation and retargeting from human data): Transferring motion from human recordings (motion capture, video) to robot kinematics and training RL policies that physically reproduce it.
    • Sterowanie całym ciałem humanoida (Humanoid whole-body control): Controlling all of a humanoid's joints at once with task priorities and contact constraints—via optimisation or a learned policy.
  • Manipulacja przedmiotami (Robotic manipulation): The branch of robotics about physical interaction with objects: pose perception, grasp selection, contact-rich motion and learning manipulation skills.
    • Douczanie robota próbami w realnym świecie z pomocą człowieka (Real-world reinforcement learning with human interventions): Reinforcement learning directly on a physical robot, initialised from demonstrations, with operator corrections during trials and reward from a success classifier.
    • Manipulacja miękkimi przedmiotami (tkaniny, liny, kable) (Deformable object manipulation): Perception, modelling and control for deformable objects (cloth, rope, cables, bags, food), whose state is very high-dimensional and whose dynamics are complex.
    • Precyzyjne zadania z kontaktem (wkładanie, montaż, wycieranie) (Contact-rich manipulation (insertion, assembly, wiping)): Manipulation requiring controlled contact and force, sensitive to small position errors: insertion, assembly, screwing, wiping, polishing.
    • Robot, który jeździ i chwyta naraz (manipulacja mobilna) (Mobile manipulation): Combining navigation of a mobile base with one- or two-arm manipulation in a single task, using whole-body or hierarchical control.
    • Wyznaczanie chwytu (Grasp detection and planning): Predicting stable gripper poses for an object from depth images or point clouds, analytically or with a learned model.
    • Zręczna manipulacja dłonią robota (Dexterous in-hand manipulation): Controlling a multi-fingered hand to change an object's pose within the grasp, including with RL policies trained in simulation and transferred to hardware.
  • Modele bazowe dla robotów (Robot foundation models): Pretrained models (VLA, LLM, VLM) used in robotics for perception, task planning and action generation across many platforms.
    • Jedna polityka dla wielu typów robotów (Cross-embodiment robot policies): Training policies on data from many platforms with different cameras, joints and action spaces, enabling transfer and quick fine-tuning.
    • Modele wizja–język–akcja (VLA) (Vision-language-action models (VLA)): Robot models combining visual perception, language instructions and action generation, developed as generalist policies transferable across tasks and robots.
      • Jak model zapisuje ruchy robota (tokeny i ciągi akcji) (Action tokenization and chunking in VLA models): Representing continuous robot motions in VLA models through discretisation, compression into tokens and prediction of trajectory chunks.
    • Planowanie zadań robota z modelem językowym (Language-model task planning for robots): Using language models to decompose instructions into sequences of robot skills or to write control code, grounded in the scene and capabilities.
  • Nawigacja robotów mobilnych (Mobile robot navigation): Localisation, mapping and path planning for mobile robots, from the classical navigation stack to learned and language-guided navigation.
    • Nawigacja według poleceń słownych (Vision-and-language navigation (VLN)): A task and family of methods in which an embodied agent navigates realistic interiors by natural-language instructions, evaluated on benchmarks such as R2R.
    • Poruszanie się robota wśród ludzi (nawigacja społeczna) (Social robot navigation): Planning mobile robot motion among pedestrians, accounting for trajectory prediction, social norms and comfort, together with evaluation methods.
    • Szukanie przedmiotu w nieznanym budynku (nawigacja do obiektu) (Object-goal navigation): An embodied navigation task: reach any object of a given category in an unseen interior using only egocentric observations.
  • Percepcja robota do działania (Robot perception for action): Action-oriented perception: estimating 6D pose, affordances and contact state from images, depth, point clouds and tactile sensors.
    • Rozpoznawanie, do czego służy przedmiot (afordancje) (Affordance learning for robot manipulation): Predicting possible actions on an object and the part that enables them (object–action–effect relation), from static data or interaction.
    • Ustalanie położenia i obrotu przedmiotu (poza 6D) (6D object pose estimation and tracking): Estimating and tracking an object's position and orientation (6 degrees of freedom) from RGB or RGB-D images, including objects unseen in training.
    • Zmysł dotyku robota (czujniki dotykowe) (Tactile and visuotactile sensing for robots): Measuring contact with tactile sensors (resistive, capacitive, visuotactile) and processing their signals into force, slip and shape.
  • Współpraca człowieka z robotem (Human-robot interaction and collaboration): Research and techniques for robots interacting with people: physical contact safety, shared control, intention communication and language interaction.
    • Bezpieczna praca ramię w ramię z robotem (koboty) (Safety in physical human-robot collaboration): Modes and methods for safe physical collaboration: monitored stop, hand guiding, speed and separation monitoring, and power and force limiting.
    • Wspólne sterowanie: człowiek prowadzi, robot pomaga (Shared autonomy and assistive teleoperation): Blending operator commands with robot autonomy under uncertainty about the user's goal, inferred online from their input history.
• Sieci i modele generatywne (Neural architectures and generative models): Model families and their mechanisms. Not every generative model is a language model.
  • Architektury efektywne (Efficient architectures): Designing a model around cost and scale.
    • Architektury hybrydowe: uwaga + warstwy liniowe (Hybrid architectures: attention plus linear/recurrent layers): Models combining full-attention layers with recurrent, state-space or linear-attention layers. Fixed-size-state layers can reduce memory cost on long sequences, while periodic full-attention layers preserve direct interactions between distant tokens; exact ratios depend on the architecture.
    • Mixture of Experts (MoE) — mieszanka ekspertów (Mixture of Experts (MoE)): Layers containing many expert subnetworks, where a router sends each token to only a selected subset. Sparse parameter activation can reduce compute per token compared with a dense model of similar total parameter count, while weight storage, communication and load balancing remain important costs.
    • Modele przestrzeni stanów (Mamba) (State-space models (Mamba)): A family of sequence models that update a fixed-size state instead of storing a full attention matrix or KV cache. Mamba introduces a selective, input-dependent state mechanism; its authors report linear scaling with sequence length and competitive results across several modalities.
    • Uwaga lokalna z przesuwnym oknem (Sliding-window (local) attention): Local attention in which a token interacts only with a bounded window of neighbouring tokens rather than the entire sequence. With a fixed window, total compute grows linearly with sequence length, and the KV memory used for generation can be bounded by the window size. Some architectures combine it with global or full attention.
  • Podstawy sieci neuronowych (Neural network fundamentals): Layers, loss functions and representation learning.
    • Funkcje aktywacji (ReLU, GELU, SwiGLU) (Activation functions (ReLU, GELU, SwiGLU)): Non-linear functions applied to neuron outputs — without them, a composition of linear layers remains a linear transformation. ReLU is common in vision networks; transformers often use GELU and gated variants such as SwiGLU in feed-forward blocks.
    • Funkcje straty (entropia krzyżowa, MSE) (Loss functions (cross-entropy, MSE)): The number training minimises: how far predictions are from the target. Cross-entropy is used for classification and next-token prediction; mean squared error (MSE) for regression and, among others, noise prediction in diffusion models.
    • Normalizacja warstw (BatchNorm, LayerNorm, RMSNorm) (Normalisation layers (BatchNorm, LayerNorm, RMSNorm)): Layers that rescale activations and often help stabilise deep-network training. BatchNorm uses mini-batch statistics and is typical in CNNs; LayerNorm normalises features within one example and is common in transformers; RMSNorm simplifies LayerNorm and is often used in newer LLMs, usually in a pre-norm layout.
    • Perceptron wielowarstwowy (MLP) (Multilayer perceptron (MLP)): A network of fully connected layers that alternate linear transformations with non-linear activation functions. The MLP is a fundamental neural model and a standard component of feed-forward blocks in transformer architectures.
    • Połączenia resztkowe (skip connections) (Residual (skip) connections): Shortcut connections that add a block's input to its output, allowing the layer to learn a correction relative to the input representation. They enabled effective training of very deep ResNets and are a common component of many modern deep architectures, including transformers.
    • Propagacja wsteczna (backpropagation) (Backpropagation): An algorithm for computing the gradient of the loss with respect to network parameters: the forward pass computes the output, while the backward pass propagates derivatives through the computation graph using the chain rule. Gradient-based optimisers use these gradients; libraries usually obtain them through automatic differentiation.
    • Regularyzacja sieci: dropout i weight decay (Neural network regularisation: dropout and weight decay): Techniques that curb overfitting in deep networks. Dropout randomly switches off units during training; weight decay penalises large weights; early stopping and data augmentation belong to the same family.
  • Reprezentacje i embeddingi (Representation learning): Encoding meaning or features as vectors.
    • Autoenkodery (ściskanie i odtwarzanie danych) (Autoencoders): An encoder–decoder network trained to reconstruct its input through a narrow hidden layer; a nonlinear PCA and the basis of VAEs and sparse autoencoders.
    • Embeddingi zdań i dokumentów (modele embeddingowe) (Sentence and document embeddings (embedding models)): Models that turn a whole sentence or document chunk into a vector compared, for example, with cosine similarity. They are commonly used in semantic retrieval and RAG pipelines, as well as clustering and deduplication; embedding models are compared on benchmarks such as MTEB. Cross-encoders for reranking and multi-vector representations complement them.
    • Osadzenia słów (word2vec, GloVe) (Word embeddings (word2vec, GloVe)): Static vectors assigned to words and learned from regularities in their contexts, so that similar uses receive nearby representations. word2vec and GloVe were important steps in the development of text representations; a single vector per word does not distinguish meanings that depend on the particular context.
    • Uczenie metryk podobieństwa (Deep metric learning): Learning an embedding space where distance reflects semantic similarity, using pairwise, triplet or proxy-based losses.
  • Rodziny modeli generatywnych (Generative model families): Different ways of learning a distribution and sampling from it.
    • Autoenkodery wariacyjne (VAE) (Variational autoencoders (VAE)): A model with an encoder that compresses data into a latent distribution and a decoder that reconstructs it from samples; trained by maximising the evidence lower bound (ELBO) with the reparameterisation trick. Today it often works behind the scenes — as the image compressor in latent diffusion.
      • Dyskretne kody obrazu i dźwięku (VQ-VAE, VQGAN) (Vector-quantised autoencoders (VQ-VAE, VQGAN)): An autoencoder whose encoder outputs indices of the nearest vectors in a learned codebook, producing discrete tokens for generative models.
      • Rozplątane reprezentacje (β-VAE) (Disentangled representations (β-VAE)): A VAE variant with weight β>1 on the KL term, strengthening independence of latent variables at the cost of reconstruction quality.
    • Autoregresyjne generowanie obrazu i dźwięku (Autoregressive image and audio generation): Models generating data as a sequence of elements (pixels, samples, tokens or scales), each conditioned on earlier ones; a rival and complement to diffusion.
      • Generowanie piksel po pikselu i próbka po próbce (PixelCNN, WaveNet) (Pixel- and sample-level autoregressive models (PixelCNN, WaveNet)): Autoregressive convolutional networks modelling pixels or audio samples sequentially using masked or dilated causal convolutions.
      • Maskowane transformery generatywne (MaskGIT) (Masked generative image transformers (MaskGIT)): A bidirectional transformer generating image tokens in parallel over a few iterations, unmasking the most confident predictions from a masked grid.
      • Przewidywanie kolejnej skali (od szkicu do detalu) (Next-scale prediction (visual autoregressive modelling, VAR)): Autoregression over scales: a transformer predicts a whole higher-resolution token map conditioned on all lower ones, instead of tokens in raster order.
    • Generatywne sieci przeciwstawne (GAN) (Generative adversarial networks (GAN)): Two networks trained in an adversarial game: a generator creates samples and a discriminator distinguishes them from real data; the discriminator's signal trains the generator to better match the data distribution. Training can be unstable and suffer from mode collapse.
      • Generator sterowany stylem (StyleGAN) (Style-based GAN generators (StyleGAN)): A GAN generator architecture where a mapping network produces a style vector injected into every synthesis layer, separating coarse attributes from fine detail.
      • Przekładanie obrazu na obraz (pix2pix, CycleGAN) (Image-to-image translation (pix2pix, CycleGAN)): Conditional GANs learning mappings between image domains, from paired examples (pix2pix) or without pairs via a cycle-consistency loss (CycleGAN).
      • Stabilne trenowanie GAN (odległość Wassersteina, kara gradientu) (Stable GAN training: Wasserstein loss and gradient penalty): A family of GAN objective changes where a critic estimates the Wasserstein distance between distributions, with the Lipschitz constraint enforced by a gradient penalty.
    • Modele dyfuzyjne (Diffusion models): Generating by gradual denoising.
      • Dyfuzja w przestrzeni pikseli (Pixel-space diffusion): Diffusion trained and sampled directly on the pixel grid, without an autoencoder.
        • Klepsydrowy transformer dyfuzyjny (Hourglass diffusion transformer): An hourglass-shaped transformer that scales linearly with pixel count.
        • Przesunięcie harmonogramu szumu (Noise-schedule shifting for high-resolution pixels): An adjustment of log signal-to-noise ratio as the pixel count grows.
        • Transformer pikselowy o dwóch poziomach (Dual-level pixel diffusion transformer): A single-stage diffusion transformer with a patch level and a pixel level.
      • Dyfuzja w przestrzeni ukrytej (Stable Diffusion) (Latent diffusion (Stable Diffusion)): Diffusion performed in an autoencoder's compressed latent space rather than directly in pixel space, reducing computation; text conditioning can be implemented through cross-attention. This design is used by Latent Diffusion models, including Stable Diffusion.
      • Dyfuzyjne modele językowe (Diffusion language models): A family of language models that learn to reverse token masking or noising, rather than only next-token prediction.
        • Dyfuzja blokowa (Block diffusion language models): An autoregressive distribution over blocks, with discrete diffusion inside a block and causal attention across blocks.
        • Dyfuzja ciągła na zanurzeniach słów (Continuous diffusion on word embeddings): Continuous diffusion in embedding space: Gaussian noise on vectors, then rounding back to a token.
        • Dyskretna dyfuzja ze score entropy (Score entropy discrete diffusion): Discrete diffusion trained with score entropy, which matches ratios of token probabilities.
        • Dyskretna dyfuzja ze strukturalną macierzą przejść (Structured discrete diffusion (D3PM)): Discrete diffusion with a transition matrix: uniform, neighbour-based, or with a mask as an absorbing state.
        • Maskowana dyfuzja tekstu (Masked diffusion language models): Diffusion whose only corruption is a mask token: the objective is a weighted mixture of masked-language losses.
          • Próbkowanie z ponownym maskowaniem (Remasking samplers for masked diffusion): Masked-diffusion samplers that can mask previously predicted tokens again and revise them over subsequent steps.
      • Kaskadowe modele dyfuzyjne (Cascaded diffusion models): A chain of diffusion models at increasing resolution, from sketch to detail.
        • Augmentacja obrazu warunkującego (Conditioning augmentation for cascaded diffusion): Corrupting the low-resolution input while training super-resolution models.
      • Sterowanie promptem bez klasyfikatora (classifier-free guidance) (Classifier-free guidance): A technique to strengthen adherence to the prompt: one model learns to generate with and without the condition (dropped randomly in training), and at sampling time the output is pushed toward the conditional version. The guidance scale is the familiar “prompt fidelity vs diversity” knob.
      • Szybkie generowanie w kilku krokach (Fast and few-step diffusion sampling): Methods for shortening diffusion sampling: better solvers without retraining and distillation into few-step and one-step models.
        • Modele spójności (generowanie jednym krokiem) (Consistency models): Models trained to map any point on a denoising trajectory to its endpoint, enabling one- and few-step sampling; trained by distillation or standalone.
        • Przeciwstawna destylacja dyfuzji (generowanie w 1–4 krokach) (Adversarial diffusion distillation): Distilling large diffusion models to 1–4 steps by combining a score-based teacher signal with an adversarial loss.
        • Stopniowa destylacja kroków (Progressive distillation of diffusion samplers): Iterative distillation where a student learns to reproduce two steps of the teacher's deterministic sampler in one, halving the step count each round.
        • Szybsze samplery bez ponownego treningu (DDIM, solvery ODE) (Fast training-free samplers (DDIM, ODE solvers)): Deterministic or less stochastic sampling procedures for trained diffusion models, based on non-Markovian processes and probability-flow ODE solvers.
      • Transformery w modelach dyfuzyjnych i przepływowych (DiT) (Transformers for diffusion and flow models (DiT)): Generative architectures in which a transformer processes patches of an image representation instead of a conventional U-Net backbone. DiT applied this design to latent diffusion, while newer variants also use transformers in flow-trained models.
      • Zaszumianie i odszumianie krok po kroku (DDPM) (Forward noising and learned denoising (DDPM)): The basic construction of denoising diffusion probabilistic models: a forward process gradually adds noise to data, while a model learns to reverse the corruption, often by predicting the added noise. Sampling-step count affects cost and quality, and separate fast-sampling or distillation techniques can reduce it.
    • Modele oparte na energii (Energy-based models (EBM)): Models defining an unnormalised density through an energy function, trained via MCMC maximum likelihood, score matching or noise-contrastive estimation.
    • Przepływy normalizujące i flow matching (Normalising flows and flow matching): Models that turn simple noise into data with an invertible transform (normalising flows) or a learned velocity field along a path from noise to data (flow matching, rectified flow). Flow matching includes diffusion as a special case and underpins many new image models, e.g. Stable Diffusion 3.
      • Prostowanie przepływu (rectified flow) (Rectified flow): A flow-matching variant learning a velocity field along straight lines between noise and data, with optional repeated straightening (reflow).
      • Przepływy odwracalne z warstwami sprzężenia (RealNVP, Glow) (Coupling-layer normalising flows (RealNVP, Glow)): Generators built from invertible transformations with tractable Jacobian determinants, giving exact log-likelihood via the change-of-variables formula.
  • Rozumowanie w modelach językowych (Reasoning in language models): Techniques and model families for tasks that require multi-step reasoning: from chain-of-thought prompting and larger inference-time compute budgets to training reasoning models with RL, SFT and distillation.
    • Efektywne rozumowanie (przeciw „overthinkingowi”) (Efficient reasoning): Methods that cut the cost and latency of reasoning: a model should not write thousands of thinking tokens for a simple question.
      • Adaptacyjne i hybrydowe tryby myślenia (Adaptive and hybrid thinking modes): One model can answer fast or think longer and decides by itself (or on request) how much reasoning a question needs.
      • Skracanie łańcucha myśli (Reasoning length reduction): Training or prompting aimed at reducing reasoning length while preserving answer quality as much as possible — for example via a length-cost signal, fine-tuning on concise traces, or brevity instructions.
    • Monitorowanie łańcucha myśli (Chain-of-thought monitoring): Analysing a model's written chain of thought (automatically or by a human) to detect signals of undesirable behaviour. Its usefulness depends in part on the monitorability and faithfulness of the reasoning trace, which are not guaranteed.
    • Obliczenia w czasie wnioskowania (Test-time compute scaling): Giving a model a bigger compute budget at answer time — without changing weights: many attempts in parallel, successive revisions, verifier-guided search, or simply thinking longer.
      • Iteracyjne poprawianie odpowiedzi (Sequential self-revision): The model reviews and revises its own answer in turns, each version building on the previous — compute grows in depth rather than width.
      • Przeszukiwanie kroków z weryfikatorem (Verifier-guided search): The solution is built step by step while a verifier (e.g. a process reward model) scores partial paths and decides which to expand.
      • Sterowanie budżetem myślenia (Thinking budget control): Explicitly setting how many tokens or how much time the model may spend reasoning — via a parameter, a cap, or forcing it to stop or continue thinking.
      • Wiele prób i głosowanie (Parallel sampling and voting): The model generates many independent solutions in parallel; the answer is picked by majority vote (self-consistency) or by a verifier score (best-of-N).
    • Pętle głębokości i myśli ukryte (Recurrent depth and latent thoughts): Scaling answer-time compute by repeating a block or a continuous state, not by extra tokens.
      • Adaptacyjne zatrzymanie obliczeń (Adaptive computation time): A differentiable rule that teaches a recurrent network how many steps to take between input and output.
      • Ciągłe myśli (Continuous chain of thought): A reasoning chain in which the hidden state returns as an embedding, without decoding to a token.
      • Głębokość rekurencyjna (Recurrent-depth language models): A language model scales answer compute by unrolling a shared recurrent block in hidden space.
      • Transformery w pętli (Looped transformers): A shared transformer block applied recurrently to sequence representations, sometimes with per-position halting.
    • Trenowanie modeli rozumujących (Training reasoning models): How to teach a model long, self-correcting reasoning: RL rewarded for correct results, SFT on long chains of thought, and distillation from a larger model.
      • Destylacja rozumowania do mniejszych modeli (Reasoning distillation): A large reasoning model generates reasoning traces that a smaller model is fine-tuned on. A cheap way to give a small model reasoning ability.
      • RL z weryfikowalnymi nagrodami (RLVR) (Reinforcement learning with verifiable rewards): The model receives reward from an automatically verifiable signal — such as answer correctness, passing code tests, or satisfying a checkable condition — rather than relying only on a learned preference model. RLVR is an important post-training approach for reasoning models, especially in domains with unambiguous verification.
        • Rekurencyjne samoulepszanie agentów terminalowych (Recursive self-improvement of terminal agents): A post-training method alternating supervised learning on verified successful trajectories with PPO using an inherited value model.
      • SFT na długich łańcuchach myśli (Long chain-of-thought fine-tuning): Supervised fine-tuning on curated examples with long reasoning traces; often a “cold start” before RL, or a cheap route to a reasoning model from few examples.
    • Weryfikatory rozumowania (Reasoning verifiers): Models or programs that judge whether reasoning is correct. Used both as a training reward and to pick the best answer at inference time.
      • Ocena każdego kroku (PRM) (Process reward models): A reward model scores each reasoning step separately, not just the final answer. Catches a mistake halfway through a solution.
      • Sprawdzanie końcowego wyniku (Outcome verification): Only the final answer is judged: by comparing with a reference, running tests, or an outcome reward model (ORM). Simpler and cheaper than step scoring.
    • Łańcuch myśli (chain-of-thought) (Chain-of-thought prompting): Getting the model to write out intermediate reasoning steps before the answer. The simplest way to improve multi-step tasks — no training required.
      • Drzewo myśli (Tree of Thoughts) (Tree of Thoughts): The model expands several alternative reasoning paths, evaluates which look promising and can backtrack — tree search instead of one line of thought.
      • Łańcuch myśli z przykładami (Few-shot chain-of-thought): The prompt includes a few example problems with step-by-step worked solutions; the model imitates that style on a new problem.
      • „Pomyśl krok po kroku” bez przykładów (Zero-shot chain-of-thought): Instead of examples, a short instruction such as “Let's think step by step” makes the model write out its reasoning.
  • Sieci grafowe (GNN) (Graph neural networks (GNN)): Neural networks operating on graphs (nodes, edges, features), invariant to node ordering; the basis of learning on molecules, social networks and meshes.
    • Konwolucje grafowe i uwaga na grafach (GCN, GAT) (Graph convolutions and graph attention (GCN, GAT)): The most common GNN layers: GCN aggregates a normalised sum of neighbour features, GAT weights neighbours with attention, GraphSAGE samples the neighbourhood.
    • Sieci geometryczne dla atomów w 3D (równoważność obrotów) (Equivariant graph networks for 3D atomic systems): GNNs over atoms with 3D positions, equivariant to rotations, translations and reflections (the E(3) group); they predict energies, forces and molecular properties.
    • Transformery grafowe (Graph transformers): Architectures combining a transformer's global attention with graph positional and structural encodings, often alongside local message passing.
  • Sieci konwolucyjne i sekwencyjne (Convolutional and recurrent networks): Architectures for spatial structure and sequences.
    • Głębokie architektury konwolucyjne (ResNet, U-Net) (Deep convolutional architectures (ResNet, U-Net)): Families of convolutional architectures: from LeNet, AlexNet and VGG through ResNet with residual connections to U-Net — an encoder-decoder with connections between corresponding levels, used for segmentation and as the backbone of some diffusion models.
    • LSTM i GRU — sieci z bramkami pamięci (LSTM and GRU (gated recurrent networks)): Recurrent networks with gating mechanisms that control how information flows into and remains in hidden state. LSTM uses a separate cell state with input, forget and output gates; GRU combines some of these mechanisms in a simpler design with reset and update gates.
    • Sieci rekurencyjne (RNN) (Recurrent neural networks (RNN)): Networks that process a sequence step by step while carrying a hidden state from the previous step. Classical RNNs can struggle to learn long dependencies because gradients vanish or explode; LSTMs and GRUs mitigate the problem, and transformers have replaced RNNs in many text tasks.
    • Splot, filtry i pooling (Convolution, filters and pooling): The basic operations of a convolutional network: small filters slide over the image and detect local patterns (edges, textures), and pooling reduces feature-map resolution. Filter size, stride and padding determine the output shape.
  • Transformery i attention (Transformers and attention): Mechanisms for mixing information across a sequence.
    • Enkoder, dekoder i enkoder–dekoder (BERT, GPT, T5) (Encoder-only, decoder-only and encoder–decoder models): The three transformer layouts. Encoders (BERT) read text bidirectionally and serve classification, search and extraction; decoders (GPT, Llama) predict the next token and make up most modern LLMs; encoder–decoders (T5, BART) map one text to another, e.g. translation and summarisation.
    • FlashAttention: Attention computed with memory in mind: less data movement, the same result.
    • Kodowanie pozycji (sinusoidalne, RoPE) (Positional encoding (sinusoidal, RoPE)): A way to provide token-order information that the attention operation alone does not encode. The original Transformer added sinusoidal vectors; many contemporary LLMs use RoPE, which rotates queries and keys according to position. For context extension RoPE is often additionally scaled or interpolated; some architectures leave part of their layers without explicit positional encoding.
    • Mechanizm self-attention (zapytania, klucze, wartości) (Self-attention (queries, keys, values)): The heart of the transformer: each token produces a query (Q), key (K) and value (V), and the output is a weighted average of values, weighted by query–key similarity across all tokens. This lets the model mix information from the whole sequence at once; cost grows quadratically with length. Text-generating models use a causal mask (a token sees only earlier ones).
    • Oszczędne warianty uwagi (MQA, GQA, MLA) (Memory-efficient attention variants (MQA, GQA, MLA)): Multi-head attention variants that reduce KV-cache memory and generation cost: MQA shares one set of keys and values across heads, GQA shares them across groups of heads, and MLA compresses key-value representations into a latent space.
• Sztuczne życie (Artificial life (ALife)): A field studying the properties and dynamics of life regardless of substrate: in software, hardware and chemistry; it includes digital evolution, cellular automata and self-organisation.
  • Automaty komórkowe (Cellular automata): Discrete models with a regular grid of finite-state cells, usually updated synchronously by a local rule depending on the neighbourhood.
    • Lenia — ciągły automat komórkowy (Lenia continuous cellular automaton): A two-dimensional cellular automaton with continuous space, time and state and a generalised local rule, producing self-sustaining “lifeforms”.
    • Neuronowe automaty komórkowe (wzory, które rosną i się naprawiają) (Neural cellular automata (NCA)): Cellular automata with continuous multi-channel state whose update rule is a neural network trained by gradient descent through many simulation steps.
  • Robotyka ewolucyjna (ewolucja sterowania i budowy robotów) (Evolutionary robotics): Applying evolutionary algorithms to design controllers, morphologies or both for physical and simulated robots, also as a tool for studying evolution.
• Trenowanie i dostrajanie (Training and adaptation): How model weights come into being and change.
  • Destylacja (Knowledge distillation): Training a model on signal from another model.
    • Destylacja kontekstu (wpisanie instrukcji w wagi modelu) (Context distillation): Fine-tuning a model so that without a long prompt it answers as it did with one.
    • Destylacja na miękkich etykietach (logity nauczyciela) (Soft-label (logit) distillation): The student learns to reproduce the teacher's full probability distribution, not just its top answer. Requires access to the teacher's outputs (logits).
    • Destylacja on-policy (uczeń generuje, nauczyciel ocenia) (On-policy distillation): The student generates its own responses and the teacher provides a dense token-level signal for them. This trains the student on states it visits itself rather than only on finished teacher-generated answers.
      • Destylacja on-policy dla hierarchicznych zespołów agentów (On-policy distillation for hierarchical multi-agent teams): Combines on-policy distillation with reinforcement learning to train a shared-parameter Leader-Searcher-Answerer team end to end.
      • Eksploracja w kontekście przez asymetryczną destylację on-policy (In-context exploration through asymmetric on-policy distillation): A student-teacher framework in which the student's on-policy history receives action labels from a privileged expert, teaching test-time exploration.
    • Destylacja polityki ze względną przewagą krokową (Step-Advantage Policy Distillation): An on-policy distillation method over recorded prefixes of multi-turn trajectories, using relative advantages among responses at a single decision step.
    • Destylacja przez odpowiedzi nauczyciela (dane syntetyczne) (Sequence-level distillation from teacher outputs): The teacher generates answers and the student is fine-tuned on them like ordinary data (SFT). Works even with an API-only teacher; licence terms need checking.
    • Rekurencyjna samodestylacja (Recursive self-distillation): Self-distillation repeated across rounds, with the updated model becoming the teacher for the next iteration.
  • Efektywne dostrajanie (Parameter-efficient fine-tuning): Training selected parameters or small adapters.
    • Adaptery (małe warstwy wstawiane do modelu) (Adapter modules): Small trainable layers are inserted into a frozen model and only they are trained. One base model can serve many tasks with separate, swappable adapters.
    • Dostrajanie przez interwencje w reprezentacje (Fine-tuning through representation interventions): Instead of changing weights, it learns small interventions on hidden representations at chosen layers and positions of a frozen model.
    • Dostrajanie wyłącznie wektorów biasu (Bias-only fine-tuning): Fine-tuning a model by updating only its bias terms or a subset of them while leaving all other weights frozen.
    • LoRA (Low-rank adaptation): Fine-tuning with low-rank adapters. Used for language and image models alike.
      • QLoRA (Quantized low-rank adaptation (QLoRA)): LoRA on top of a base model stored in 4 bits. Cuts the GPU memory needed for fine-tuning.
      • Warianty LoRA (DoRA, LoRA+, AdaLoRA) (LoRA variants): Refinements of basic LoRA: separating weight magnitude and direction, different learning rates for the two adapter matrices, adaptive rank allocation. They aim to close the gap to full fine-tuning.
    • Miękkie prompty (prompt tuning, prefix tuning) (Soft prompts: prompt and prefix tuning): The model stays frozen while a small set of continuous vectors is learned. Prompt tuning adds them as virtual input tokens, whereas prefix tuning supplies learned prefixes to successive attention layers.
    • Skalowanie aktywacji wyuczonymi wektorami (Learned-vector activation rescaling): A PEFT method that multiplies attention keys, values and feed-forward activations by learned vectors.
  • Post-training: Shaping a pretrained model toward the behaviour you want.
    • Preferencje i RL dla LLM (Preference optimization and RL for LLMs): Shaping behaviour with preferences or rewards.
      • DPO i optymalizacja offline (Direct preference optimization and offline variants): Training directly on chosen–rejected answer pairs, with no separate reward model. Computationally cheaper than RL.
        • Iteracyjne DPO na świeżych danych preferencji (Iterative DPO with newly generated preference data): Repeated rounds of DPO in which the current model generates answer pairs and a human, reward model or judge rates them.
        • Optymalizacja preferencji bez modelu referencyjnego (Reference-free preference optimization): DPO variants that need no frozen copy of the model as a reference point.
      • Modele nagrody (Reward models): A model trained on human judgements that scores answers. Used for RL, data filtering and evaluation.
      • Optymalizacja preferencji z ocen binarnych (Preference optimization from binary desirability feedback): Aligning a model on single responses labelled desirable or undesirable, using a utility function from prospect theory.
      • RL z nagrodą (Reinforcement learning with rewards): Iteratively improving the model's policy against a reward while staying close to its starting behaviour. Costlier and sensitive to hyperparameters.
        • Destylacja posterioru świadoma drzewa dla długich agentów (Tree-aware posterior self-distillation for long-horizon agentic RL): Critic-free tree-aware posterior self-distillation for credit assignment in agentic RL.
        • GRPO i warianty bez krytyka (GRPO and critic-free policy optimization): A PPO variant without a separate value model: the model samples a group of answers to the same prompt and estimates advantage relative to rewards within that group. This reduces memory requirements and is used in several newer reasoning-training recipes.
        • Klasyczny RLHF: SFT → model nagrody → PPO (Classic RLHF pipeline with PPO): The three-stage RLHF recipe popularised by InstructGPT: supervised fine-tuning on demonstrations, training a reward model from human rankings, then policy optimisation with PPO while constraining divergence from a reference model.
        • Uczenie agentów na trajektoriach wielosesyjnych (Multi-session agentic reinforcement learning): Critic-free hierarchical advantage optimization for agent trajectories composed of multiple sessions.
    • SFT / instruction tuning (Supervised fine-tuning): Fine-tuning on examples of the expected answer.
      • Dostrajanie do wywoływania funkcji i narzędzi (Function-calling and tool-use fine-tuning): Fine-tuning a model on examples whose answer is a correct, structured tool call with arguments.
    • Szablony czatu i formaty rozmów (Chat templates and conversation formats): How a conversation is stored in a dataset and turned into text for the model. The training template has to match at inference.
  • Proces uczenia (Training optimization): Steering how the weights get updated.
    • Harmonogram learning rate (rozgrzewka, cosinus, WSD) (Learning-rate schedules): A plan for changing the learning rate during training. Common patterns include warm-up followed by decay; WSD adds a longer stable-rate phase before the final decay, making training runs easier to extend.
    • Optymalizatory (SGD, Adam/AdamW, Muon) (Optimizers for neural network training): Rules that turn a gradient into a weight update: from SGD with momentum, through adaptive Adam and AdamW, to newer matrix-aware optimisers such as Muon.
    • Oszczędzanie pamięci podczas treningu (Memory-efficient training): Techniques that shrink memory used by activations, gradients and optimiser states, so larger models train on the same hardware.
      • 8-bitowe stany optymalizatora (8-bit optimizer states): Storing optimiser statistics (e.g. Adam moments) in 8 bits with block-wise quantisation instead of 32 bits.
      • Niskorzędowe rzutowanie gradientu (Low-rank gradient projection): Full-parameter training in which gradients and optimiser states are kept in a low-dimensional projection.
      • Ponowne liczenie aktywacji (gradient checkpointing) (Gradient (activation) checkpointing): Only selected activations are stored; the missing ones are recomputed during backpropagation.
    • Pakowanie sekwencji w treningu (sequence packing) (Sequence packing for training): Concatenating several short examples into one context-length sequence to reduce padding in a batch.
    • Przenoszenie hiperparametrów z małego modelu na duży (μP) (Maximal update parametrization (μP) and hyperparameter transfer): A network parametrisation under which optimal hyperparameters are tuned on a small model and transferred unchanged to a large one.
    • Rozmiar partii treningowej i akumulacja gradientu (Training batch size and gradient accumulation): Choosing how many examples go into each update step: gradient accumulation over micro-batches and the critical batch size.
    • Stabilność treningu i skoki straty (Training stability): Diagnosing and mitigating instabilities such as sudden spikes or divergence in loss during long training runs of large models. The problems can be reproduced in smaller models to test safeguards in the training recipe.
    • Trening w mieszanej precyzji (FP16, BF16, FP8) (Mixed-precision training): Running most operations in shorter number formats while keeping key values in full precision. It cuts memory use and speeds up training on modern GPUs.
  • Trening od zera / pretraining (Training from scratch and pretraining): Teaching a model from initial weights. Pretraining is the stage that prepares it for later use or adaptation.
    • Maskowane modelowanie języka (MLM) (Masked language modelling (MLM)): A pretraining objective for encoders: the model restores masked tokens from context on both sides.
    • Mid-training (późna faza pretrainingu) (Mid-training): A late pretraining phase on a smaller, carefully chosen data mix (e.g. maths, code, reasoning, long documents), usually while the learning rate is annealed. It strengthens selected capabilities before post-training.
    • Prawa skalowania i budżet obliczeń (Scaling laws and compute-optimal training): Predictable relationships between model size, data, compute and quality. They let you plan a large training run from small ones and choose the ratio of parameters to tokens.
    • Przewidywanie następnego tokenu (modele autoregresyjne) (Next-token prediction objective): A common pretraining objective for autoregressive language models: predict the next token from the preceding ones. The error is typically optimised as cross-entropy.
    • Przewidywanie wielu tokenów naraz (multi-token prediction) (Multi-token prediction): An auxiliary pretraining objective: from each position the model predicts several upcoming tokens with separate heads.
    • Wydłużanie okna kontekstu (Context window extension): Further training intended to improve performance on sequences longer than those used previously — often by modifying positional encoding and training on long examples.
  • Uczenie ciągłe i transfer (Continual and transfer learning): Adapting to new data and new tasks.
    • Dotrenowanie na tekstach z dziedziny (continued pretraining) (Continued (domain-adaptive) pretraining): An additional pretraining phase of an existing model on unlabelled data from a selected domain, task or language, usually before downstream fine-tuning. It aims to adapt the model to a new data distribution and vocabulary.
    • Edycja wiedzy w modelu (ROME, MEMIT) (Knowledge editing in language models): A targeted change to facts or behaviours stored in a model without full retraining — by editing weights or using external memory. The goal is to limit side effects outside the edit's scope, but this is not guaranteed.
    • Katastrofalne zapominanie i jego ograniczanie (Catastrophic forgetting and its mitigation): While learning new data or tasks, a network may lose earlier capabilities. Mitigation methods include replaying earlier data, protecting important weights and reducing the scope of parameter updates.
    • Selektywny nadzór eksperta w ciągłym treningu agentów (Selective expert supervision in continual agent reinforcement learning): RL as backbone with selective SFT activated on hard states for continual agent training.
  • Uczenie w czasie testu (Test-time training): Updating part of the parameters on an unlabeled test sample before the actual prediction.
    • Maskowanie obrazu w czasie testu (Masked autoencoder test-time training): One-sample test-time training whose auxiliary objective is reconstructing a masked image.
    • Warstwy z modelem w stanie ukrytym (Test-time training layers): A sequence layer whose hidden state is a model updated by a self-supervised step also at test time.
    • Zadanie pomocnicze na jednej próbce (One-sample self-supervised test-time training): A short self-supervised update on one sample or a stream, before the main-task prediction.
  • Łączenie modeli (Model merging): Building a new model by arithmetic on the weights of already-trained models, with no further training.
    • Arytmetyka zadań (task vectors, TIES, DARE) (Task arithmetic and task vectors): The difference between fine-tuned and base-model weights forms a 'task vector' that can be added, subtracted and combined. TIES resolves sign conflicts between vectors, while DARE sparsifies delta parameters before merging.
    • Ewolucyjne łączenie modeli (Evolutionary model merging): An evolutionary algorithm searches weight mixtures or layer compositions instead of relying on one hand-designed recipe. The source paper also demonstrates merging models specialised in different domains.
    • Uśrednianie i interpolacja wag (model soups, SLERP) (Weight averaging and interpolation): Combining several fine-tunes of the same base model by averaging or smoothly interpolating their weights. Often improves quality and robustness at no extra inference cost.
• Uczenie maszynowe (Machine learning): Core learning tasks and methods. Deep learning sits in its own branch for readability.
  • Drzewa i metody zespołowe (Tree ensembles): Combining simpler models into a stronger predictor.
    • Drzewa decyzyjne (Decision trees): A model that asks a series of if–then questions about features until it reaches an answer. Easy to read but a single tree overfits easily; it is the building block of random forests and boosting.
    • Gradient boosting (XGBoost, LightGBM) (Gradient-boosted trees): Trees added one after another, each correcting the errors of the previous ones. For years the standard for tabular data and competitions.
    • Las losowy i bagging (Random forests and bagging): Many trees trained on random samples of rows and features, with their votes averaged. A robust out-of-the-box model that is not very sensitive to settings.
    • Stacking i łączenie modeli (Stacking and model blending): Stacked generalisation: a meta-model trained on out-of-fold predictions of several base models.
  • Prognozy i anomalie (Forecasting and anomaly detection): Predicting change and spotting unusual events.
    • Klasyczne prognozowanie szeregów czasowych (Classical time-series forecasting): Statistical models of trend, seasonality and dependence on past values; still a strong baseline for business forecasting.
    • Modele bazowe dla szeregów czasowych (Time-series foundation models): Transformers pretrained on a very large number of diverse series that forecast a new series without training on it (zero-shot).
    • Wykrywanie anomalii i nowości (Outlier and novelty detection): Finding examples that do not fit the rest of the data (outliers) or what the model has seen before (novelties), e.g. fraud, failures or unusual readings.
  • Przygotowanie cech (feature engineering) (Feature engineering and preprocessing): Turning raw columns into features a model can learn from: scaling, encoding categories, imputing missing values, creating and selecting features. Steps are chained in a pipeline so no test-set information leaks in.
    • Kodowanie zmiennych kategorycznych (Categorical feature encoding): Converting categorical features into numeric representations: one-hot, ordinal, target encoding or learned embeddings.
    • Selekcja cech (Feature selection): Choosing a feature subset with filter methods (tests, mutual information), wrapper methods (RFE, sequential selection) and embedded methods (L1, feature importance).
    • Skalowanie i normalizacja cech (Feature scaling and normalisation): Transformations of numeric features: standardisation, min–max, robust scaling, and log, power and quantile transforms.
    • Uzupełnianie brakujących danych (imputacja) (Missing data imputation): Replacing missing entries with estimates from simple statistics, neighbours or a model of other features, often with an added missingness indicator.
  • Systemy rekomendacyjne (Recommendation systems): Matching content or products to a person and context.
    • Filtrowanie kolaboracyjne i faktoryzacja macierzy (Collaborative filtering and matrix factorisation): Recommending from the behaviour of similar people (“who bought X also bought Y”); matrix factorisation learns user and item embeddings.
    • Generowanie kandydatów, scoring i re-ranking (Candidate generation, scoring and re-ranking): The typical architecture of large systems: quickly narrow a huge catalogue to hundreds of candidates, score them more precisely, then account for freshness, diversity and business rules.
    • Rekomendacje oparte na treści (Content-based filtering): Recommending items similar to what a person already likes, based on features of the items themselves.
  • Uczenie nadzorowane (Supervised learning): Learning from examples with the answer attached.
    • Maszyny wektorów nośnych (SVM) (Support vector machines): A classifier that looks for the boundary with the widest margin between classes; kernels (the kernel trick) let it draw non-linear boundaries.
    • Metoda k najbliższych sąsiadów (k-NN) (k-nearest neighbours): A model with no parameter training: the answer for a new example comes from the k most similar examples in the data. The same idea underlies vector search.
    • Modele bazowe dla danych tabelarycznych (Tabular foundation models): Transformers pretrained on millions of (mostly synthetic) tables that predict on a new table without training, learning in context from the given rows. A new direction (since 2025) competing with gradient boosting on small and medium datasets.
    • Naiwny klasyfikator Bayesa (Naive Bayes classifiers): A probabilistic classifier built on Bayes' theorem with the simplifying assumption that features are independent. A fast baseline, classically for text (e.g. spam filtering).
    • Regresja kwantylowa i przedziały predykcji (Quantile regression and prediction intervals): Estimating conditional quantiles of the target by minimising the pinball loss; a pair of quantiles forms a prediction interval.
    • Regresja liniowa i logistyczna (Linear and logistic regression): The simplest predictive models: a weighted sum of features gives a number (linear regression) or a class probability (logistic regression). They are the usual place to learn loss, gradient descent and decision thresholds.
    • Uczenie rankingu (learning to rank) (Learning to rank): Learning a scoring function for documents given a query, with pointwise, pairwise or listwise objectives that optimise ranking-quality metrics.
  • Uczenie nienadzorowane (Unsupervised learning): Finding structure without ready labels.
    • Klasteryzacja (Clustering): Grouping similar examples without labels, e.g. customer segmentation or organising documents.
      • Klasteryzacja gęstościowa (DBSCAN, HDBSCAN) (Density-based clustering): Clusters are dense regions of data; points in sparse areas are marked as noise. No need to set the number of clusters, and it handles irregular shapes.
      • Klasteryzacja hierarchiczna (Hierarchical clustering): Builds a tree of groups by gradually merging the closest ones (or splitting the largest); the cut level is chosen from a dendrogram.
      • Metoda k-średnich (k-means) (k-means clustering): Splits data into a preset number k of groups around centroids that are moved iteratively. Works well for compact groups of similar size.
    • Modelowanie tematów (Topic modelling): Unsupervised discovery of themes in a corpus: probabilistic word mixtures (LDA) or clustering of embeddings with topic descriptions (BERTopic).
    • Redukcja wymiarowości (Dimensionality reduction): Turning many features into a few that keep the most important structure of the data — for visualisation, compression or as a step before another model.
      • Analiza głównych składowych (PCA) (Principal component analysis): A linear method that finds the directions of largest variance in the data and projects the data onto them.
      • Mapy danych w 2D: t-SNE i UMAP (Nonlinear embedding for visualisation (t-SNE, UMAP)): Non-linear methods that lay points out in 2D so that neighbours stay close. Popular for inspecting embeddings and clusters; distances between groups on such maps are easy to over-interpret.
    • Reguły asocjacyjne i częste zbiory (Association rules and frequent itemsets): Mining frequent itemsets and X→Y rules in transactional data, scored by support, confidence and lift.
  • Uczenie przyrostowe (online) na strumieniach danych (Online and incremental learning on data streams): Updating a model after single samples or small chunks of a stream, with predict-then-learn evaluation and reaction to concept drift.
  • Uczenie wielozadaniowe (Multi-task learning): Joint training on several tasks with a shared representation, acting as regularisation and knowledge transfer between tasks.
  • Uczenie z ograniczonym lub automatycznym nadzorem (Learning with limited or automatically derived supervision): Learning paradigms where supervision is partial, generated automatically from data, or attached to groups rather than individual examples.
    • Uczenie pół-nadzorowane (Semi-supervised learning): Combining a small labelled set with a large unlabelled one through pseudo-labels, consistency regularisation or label propagation over a similarity graph.
      • Samouczenie i pseudoetykiety (Self-training and pseudo-labelling): Iteratively adding the model's above-threshold predictions to training as labels; modern variants combine it with augmentation and consistency.
    • Uczenie samonadzorowane (Self-supervised learning): Learning representations from a signal generated by the data itself (pretext tasks, contrast, masking), later transferred to downstream tasks.
    • Uczenie wieloinstancyjne (MIL) (Multiple instance learning (MIL)): Learning from labels attached to bags of instances; the model aggregates instance representations, e.g. with attention, into a bag-level prediction.
  • Walidacja i dobór modelu (Model selection and validation): How to check reliably whether a model generalises to new data, and how to choose a model and its settings without fooling yourself.
    • AutoML — automatyczny dobór modeli (Automated machine learning (AutoML)): Tools that prepare features, try many models and settings, and ensemble the best ones automatically.
    • Przeuczenie, niedouczenie i regularyzacja (Overfitting, underfitting and regularisation): The trade-off between a model that is too simple (misses the pattern) and one fitted too closely to training data (learns noise), and techniques that counter it.
    • Strojenie hiperparametrów (Hyperparameter tuning): Searching for model settings that are not learned from data (tree depth, regularisation strength, learning rate) using validation results.
    • Walidacja krzyżowa (Cross-validation): Repeatedly splitting data into training and validation parts (e.g. 5 folds) so the estimate does not depend on one lucky split.
• Uczenie ze wzmocnieniem i decyzje (Reinforcement learning and sequential decision-making): Learning to act from outcomes and interaction with an environment.
  • Algorytmy RL (Reinforcement learning algorithms): Methods for learning values and policies.
    • Bezpieczne RL i RL z ograniczeniami (Safe and constrained reinforcement learning): Learning policies that maximise reward while aiming to keep cost or risk within set limits during training or deployment, e.g. through a separate cap on expected cost instead of folding penalties into the reward. Guarantees usually hold in expectation, not for every run.
    • Bufor doświadczeń i priorytetowe odtwarzanie (Experience replay and prioritised replay): A mechanism that stores transitions for repeated off-policy learning; prioritised replay samples more informative experiences more often.
    • Deterministyczne gradienty polityki (Deterministic policy-gradient methods): Off-policy actor–critic methods for continuous actions with a deterministic actor; TD3 adds fixes that curb DDPG's value overestimation.
    • Gradient polityki (REINFORCE) (Policy gradient methods (REINFORCE)): Directly adjusting policy parameters towards actions that led to higher reward. Simple and general, but sample-hungry.
    • Hierarchiczne RL i opcje (Hierarchical reinforcement learning and options): Splitting behaviour into levels: a higher-level policy picks multi-step skills (options), which in turn choose low-level moves and when to stop.
    • Metody aktor–krytyk (A2C, SAC) (Actor–critic methods (A2C, SAC)): Combines a learned policy (actor) with a value or Q function (critic) that guides policy updates and can reduce gradient-estimator variance. A widely used family for continuous control.
    • Proximal Policy Optimization (PPO) (Proximal policy optimization (PPO)): A policy-gradient method using a surrogate objective designed to constrain overly large policy changes during updates. Widely used in simulation, robotics and some RLHF systems.
    • Q-learning i głębokie sieci Q (DQN) (Q-learning and deep Q-networks (DQN)): Off-policy learning of action values from the Bellman equation; in DQN a neural network replaces the Q-table. Methods in this family reuse data from past interactions, but training with function approximation needs stabilising mechanisms.
    • Rozkładowe uczenie ze wzmocnieniem (Distributional reinforcement learning): A family of methods learning the distribution of returns (value distribution) instead of its expectation, via a distributional Bellman equation.
    • Uczenie z celami i odtwarzanie z perspektywy czasu (HER) (Goal-conditioned RL and hindsight experience replay): One policy trained for many goals at once, with the goal given as an extra input. Hindsight replay relabels failed attempts as if the goal had been whatever was actually reached, enabling learning from success-only rewards.
  • Modele świata i planowanie (Model-based reinforcement learning): Predicting the effect of actions before taking them.
    • Architektury predykcyjne wspólnego osadzenia (JEPA) (Joint Embedding Predictive Architectures (JEPA)): Models that learn to predict abstract representations of future or missing states in embedding space, rather than reconstructing raw data (pixels, tokens).
      • Hierarchiczne JEPA (H-JEPA) (Hierarchical Joint Embedding Predictive Architectures (H-JEPA)): An extension of JEPA with planning at multiple levels of abstraction, using separate representations for different timescales.
      • JEPA-Anything – uniwersalne predykcyjne modele świata (JEPA-Anything: domain-agnostic predictive world models): An extension of JEPA with orthogonal predictive factorization (OPF), enabling a shared learning principle for world models across many different domains.
    • Generatywne modele świata dla agentów i robotów (Generative world models for agents and robotics): Models that generate or predict how an environment evolves, sometimes conditioned on user or agent actions. They can create interactive environments or provide simulation foundations for training and testing autonomous systems.
    • Planowanie z przeszukiwaniem drzewa (AlphaZero, MuZero) (Tree-search planning with learned models): Combining Monte Carlo tree search with networks that predict value and policy, and in MuZero also a learned model of the rules. Key to the successes in board games.
    • Uczenie w wyobraźni modelu świata (Dreamer) (Learning in imagination with latent world models): The agent learns a compact model of the environment and trains its policy on imagined rollouts instead of costly real trials.
  • Offline RL i eksploracja (Offline reinforcement learning and exploration): Learning from a fixed set of experience, and gathering new experience.
    • Eksploracja z ciekawości (nagroda wewnętrzna) (Curiosity-driven exploration and intrinsic rewards): The agent earns an extra reward for novel or hard-to-predict situations, which helps when external reward is sparse.
    • Konserwatywne offline RL (CQL, IQL) (Conservative offline RL): Learning a policy from a fixed dataset without new interaction; penalising overestimated out-of-dataset actions guards against distribution-shift errors.
    • Ocena polityki off-policy (OPE) (Off-policy evaluation (OPE)): Estimating how good a new policy would be using only data collected by a different policy, without running it. Often a critical step before deploying RL in real systems.
    • RL jako modelowanie sekwencji (Decision Transformer) (RL as sequence modelling (Decision Transformer)): Treating trajectories like text: a transformer predicts the next action from history and a desired return.
  • Podstawy RL (Reinforcement learning fundamentals): States, actions, rewards and policies.
    • Częściowa obserwowalność (POMDP) (Partially observable MDPs (POMDP)): Decision-making when observations do not reveal the full state of the world, so the agent must combine information over time, e.g. through a belief state or network memory.
    • Funkcje wartości i równanie Bellmana (Value functions and the Bellman equation): How much reward to expect from a state or action; the Bellman equation links the value now to the value one step later. The basis of Q-learning and actor–critic methods.
    • Proces decyzyjny Markowa (MDP) (Markov decision processes (MDP)): A formal model of sequential decision-making in RL: states, actions, transition probabilities, rewards and a discount factor. One of the standard starting points for describing RL algorithms.
    • Programowanie dynamiczne: iteracja wartości i polityki (Dynamic programming for MDPs (value and policy iteration)): Computing an optimal policy when the transition and reward model is fully known, by repeatedly applying the Bellman equations. The reference point for methods that learn from trials.
    • Projektowanie i kształtowanie nagrody (Reward design and reward shaping): Specifying the reward signal so the agent learns the intended behaviour, and adding auxiliary rewards that speed up learning without changing the goal.
    • Uczenie z różnic czasowych (TD) i metody Monte Carlo (Temporal-difference learning and Monte Carlo methods): Learning value functions from experience without an environment model: Monte Carlo averages full episode returns, while TD updates estimates after every step using the next state's estimate.
    • Wielorękie bandyty: eksploracja kontra eksploatacja (Multi-armed and contextual bandits): Repeatedly choosing among options with unknown rewards while balancing exploration of new possibilities against exploitation of the current best. Used in recommendation and adaptive online experimentation.
    • Środowiska i symulatory do RL (RL environments and simulators): The standard reset/step interface and ready-made environments where agents practise; also physics simulators for robots.
    • „Śmiertelna triada” w RL (The deadly triad in reinforcement learning): Generalising value functions with approximators, and the instability arising from combining approximation, bootstrapping and off-policy learning.
  • Uczenie przez imitację (Imitation learning): Learning to act from demonstrations.
    • Imitacja z korektą eksperta (DAgger) (Interactive imitation learning (DAgger)): An expert labels correct actions in states reached by the learning policy, and the new examples are added to the dataset. This mitigates compounding error in behavioural cloning.
    • Klonowanie zachowań (behavioral cloning) (Behavioural cloning): Supervised learning on observation→action pairs from demonstrations. It is a foundational way to learn robot policies from demonstrations; a weakness is compounding error outside the distribution represented in the data.
    • Odwrotne RL i imitacja adwersarialna (IRL, GAIL) (Inverse RL and adversarial imitation): Instead of copying actions, recover the goal (reward) that explains expert behaviour, or train a policy so a discriminator cannot tell it from the demonstrations.
    • Polityki generatywne (Diffusion Policy, ACT) (Generative action policies (Diffusion Policy, ACT)): The policy generates whole chunks of future actions with a generative model (diffusion, flow matching, transformer), capturing the different ways demonstrators perform a task.
  • Uczenie wieloagentowe (MARL) (Multi-agent reinforcement learning (MARL)): Many learning agents in one environment: cooperation, competition and self-play. Combines RL with game theory.
    • Scentralizowany trening, zdecentralizowane działanie (CTDE) (Centralised training for decentralised execution (CTDE)): A cooperative MARL paradigm: training may use global state and other agents' data, while execution relies only on local observation histories.
      • Faktoryzacja wartości w kooperacyjnym MARL (Value-function factorisation for cooperative MARL): Value-based CTDE methods representing the joint Q-function as a composition of per-agent Q-functions consistent with the Individual-Global-Max principle.
      • Scentralizowani krytycy w wieloagentowym aktorze–krytyku (Centralised critics in multi-agent actor–critic): Multi-agent actor–critic methods where decentralised actors learn from critics that can access additional information about other agents during training.
    • Uczenie komunikacji między agentami (Learned multi-agent communication (emergent communication)): A MARL direction in which partially observing agents learn communication protocols end-to-end, together with their action policies.
    • Współpraca z nieznanymi partnerami (ad hoc teamwork) (Ad hoc teamwork): The problem of designing agents that cooperate with previously unseen partners without prior coordination, including zero-shot coordination with humans.
  • Uogólnianie i adaptacja w RL (Generalisation and adaptation in reinforcement learning): A family of RL methods concerned with transferring policies to new environments, rapid task adaptation and learning under nonstationarity.
    • Augmentacja obserwacji w RL z obrazu (Data augmentation for pixel-based reinforcement learning): Regularising RL from visual observations through random transformations of replay-buffer images, improving sample efficiency and robustness.
    • Automatyczne projektowanie środowisk treningowych (UED) (Unsupervised environment design for reinforcement learning): Methods that automatically generate or select environment configurations to form a curriculum that improves agent robustness and generalisation.
    • Ciągłe uczenie ze wzmocnieniem (Continual reinforcement learning): Reinforcement learning under nonstationarity without a fixed end of training, including the loss of plasticity in deep networks.
    • Meta-uczenie ze wzmocnieniem (meta-RL) (Meta-reinforcement learning): Meta-learning applied to RL: training over a task distribution yields a policy or algorithm that adapts to a new task from few interactions.