1"""Area ai_ml, part B: categories 5-8 and 38 (agents, RL, ML theory, DL architectures, AI safety)."""23CATEGORIES = [4dict(5 id="ai_agents", number=5, name="AI Agents and Autonomous Systems", area="ai_ml",6 definition=("Software agents that perceive, decide and act with some autonomy: LLM-based agents and tool use, "7 "multi-agent systems, planning and automated reasoning for acting, agent architectures, negotiation, "8 "agentic workflows and benchmarks, embodied agents in simulated environments. Physical robots are "9 "robotics; RL algorithms as such are reinforcement_learning."),10 topics=[("T10456", 1.0, "Multi-Agent Systems and Negotiation: BDI, agent architectures, argumentation"),11 ("T10906", 0.6, "AI-based problem solving and planning (PDDL, multi-agent planning)"),12 ("T11574", 0.4, "AI in games (game-playing agents)"),13 ("T10249", 0.3, "distributed control of multi-agent systems (control-theoretic MAS)"),14 ("T12128", 0.2, "chatbots and service agents")],15 subfields=[(1702, 0.1)],16 arxiv=[("cs.MA", 1.0), ("cs.AI", 0.4)],17 streams=[("conf/atal", 1.0), ("conf/ifaamas", 1.0), ("journals/aamas", 1.0), ("conf/icaart", 0.6),18 ("conf/aaai", 0.2), ("conf/ijcai", 0.2), ("journals/jair", 0.2), ("journals/ai", 0.2), ("conf/kr", 0.2),19 ("conf/ecai", 0.3), ("conf/ecai2", 0.3), ("conf/ictai", 0.2)],20 keywords=[21 (r"(?i)\b(LLM|language[- ]model|AI|autonomous|intelligent|embodied|conversational|software|web|GUI|coding|browsing|tool-?using|generative|multi-?step|computer-?use|research|planning)[- ](based |powered |driven )?agents?\b", 1.0, "both"),22 (r"(?i)\bagentic\b|\bagent-?based (LLM|AI|systems?|frameworks?|workflows?)\b", 1.0, "both"),23 (r"(?i)\bmulti-?agent (systems?|reinforcement|collaboration|coordination|negotiation|planning|simulation|frameworks?|debate|learning|path finding|communication|environments?)\b|\bMAS\b|\bMARL\b", 1.0, "both"),24 (r"(?i)\b(tool (use|using|learning|calling|augmented|integration)|function calling|API calling|ReAct|AutoGPT|AutoGen|LangChain|LangGraph|CrewAI|MetaGPT|SWE-?agent|SWE-?bench|WebArena|AgentBench|OSWorld|Model Context Protocol|MCP servers?)\b", 1.0, "both"),25 (r"(?i)\b(BDI|belief-?desire-?intention) (agents?|architectures?|models?)\b|\bagent-?oriented (programming|software engineering)\b", 1.0, "both"),26 (r"(?i)\b(automated|classical|hierarchical|task|AI|temporal|epistemic) planning\b|\bPDDL\b|\btask (decomposition|and motion planning)\b", 0.7, "both"),27 (r"(?i)\bagent-?based (model(ing|s)?|simulations?)\b|\bABM\b", 0.6, "both"),28 (r"(?i)\b(autonomous systems?|autonomy|self-?driving|self-?improving|self-?evolving|long-?horizon (tasks?|planning)|memory (module|mechanism|augmented) (for|in) agents?)\b", 0.5, "both"),29 (r"(?i)\b(game[- ]playing|general game|AlphaGo|AlphaZero|MuZero|OpenAI Five|AlphaStar|CICERO|Diplomacy)\b", 0.6, "both"),30 ],31 exclusions=[r"(?i)\b(reagents?|chemical|contrast|therapeutic|antimicrobial|chelating|reducing|oxidi[sz]ing|pharmacological|anti-?cancer|infectious|causative|etiologic(al)?|biological|pathogenic|anesthetic|anaesthetic|alkylating|cross-?linking|blowing|surface-?active|wetting|coupling|cytotoxic|antiviral|antibacterial|antifungal|bleaching|dispersing|sweetening|flavou?ring|leavening|real estate|travel|insurance|customs|shipping|sports|literary|secret|double|free|special) agents?\b"],32 related=["llm", "reinforcement_learning", "robotics", "ai_safety", "autonomous_scientific_discovery"],33 parent_of=[], child_of=[],34 provenance=("T10456 is the classical MAS host; the LLM-agent wave has no OpenAlex topic yet, so keyword patterns on "35 "'X agent' (with the qualifier restricted to computing senses) and agent frameworks/benchmarks carry it. "36 "Chemical/medical/commercial senses of 'agent' are vetoed."),37),38dict(39 id="reinforcement_learning", number=6, name="Reinforcement Learning", area="ai_ml",40 definition=("Learning to act from reward signals: value- and policy-based methods, model-based RL, offline and "41 "inverse RL, multi-agent RL, bandits, exploration, reward design, RL from human feedback, and RL theory. "42 "Applications to robot control are shared with robotics."),43 topics=[("T10462", 1.0, "Reinforcement Learning in Robotics: Q-learning, PPO, offline RL, imitation learning"),44 ("T12794", 0.5, "adaptive dynamic programming control"),45 ("T11574", 0.3, "AI in games"),46 ("T10100", 0.1, "metaheuristic optimisation (RL-based hyper-heuristics)")],47 subfields=[],48 arxiv=[("cs.LG", 0.3), ("cs.AI", 0.2), ("cs.RO", 0.1), ("eess.SY", 0.1), ("cs.MA", 0.1)],49 streams=[("conf/nips", 0.2), ("conf/icml", 0.2), ("conf/iclr", 0.2), ("conf/corl", 0.3), ("conf/aaai", 0.1),50 ("journals/jmlr", 0.1), ("conf/atal", 0.2), ("journals/tnn", 0.1)],51 keywords=[52 (r"(?i)\breinforcement learning\b", 1.0, "both"),53 (r"(?i)\b(deep|multi-?agent|offline|online|inverse|model-?based|model-?free|safe|hierarchical|meta|batch|distributional|constrained|goal-?conditioned|preference-?based|federated|quantum)[- ]RL\b", 1.0, "both"),54 (r"(?i)\b(Q-?learning|DQN|double DQN|Rainbow DQN|PPO|proximal policy optimi[sz]ation|TRPO|actor-?critic|A3C|A2C|SAC\b|soft actor-?critic|TD3|DDPG|policy gradients?|REINFORCE|AlphaZero|MuZero|Dreamer(V\d)?|decision transformers?|world models? for (control|RL|planning))\b", 1.0, "both"),55 (r"(?i)\bMarkov decision process(es)?\b|\bMDPs?\b|\bPOMDPs?\b|\bsemi-?MDP\b|\bDec-?POMDP\b", 0.8, "both"),56 (r"(?i)\b(reward (functions?|shaping|models?|hacking|design|learning|signals?)|sparse rewards?|exploration[- ]exploitation|intrinsic (motivation|rewards?)|curiosity-?driven|multi-?armed bandits?|contextual bandits?|credit assignment|experience replay|sim-?to-?real)\b", 0.8, "both"),57 (r"(?i)\b(imitation learning|behavio(u)?r(al)? cloning|learning from demonstrations?|inverse reinforcement|apprenticeship learning|offline (policy|data) (evaluation|learning))\b", 0.8, "both"),58 (r"(?i)\b(RLHF|RLAIF|reinforcement learning from (human|AI) feedback|GRPO|DPO\b|direct preference optimi[sz]ation|RLVR|reasoning models? trained with|verifiable rewards?)\b", 0.8, "both"),59 (r"(?i)\b(temporal[- ]difference|value (functions?|iteration)|policy (iteration|evaluation|optimi[sz]ation|learning)|Bellman|dynamic programming for control|optimal control via learning|regret bounds? for (MDP|RL))\b", 0.6, "both"),60 (r"(?i)\b(agents?|policies|policy|environments?|episodes?|rollouts?)\b", 0.2, "both"),61 ],62 exclusions=[r"(?i)\breinforcement (of|in|for) (concrete|steel|soil|composites?|polymers?|fib(er|re)s?|masonry|beams?|slabs?|pavements?)\b(?!.*reinforcement learning)"],63 related=["ai_agents", "robotics", "ml_theory", "ai_safety", "llm"],64 parent_of=[], child_of=[],65 provenance="T10462 is the RL topic (its keywords list MARL, Q-learning, PPO, offline RL). Algorithm names reach the threshold alone.",66),67dict(68 id="ml_theory", number=7, name="Machine Learning Theory", area="ai_ml",69 definition=("Statistical and computational learning theory and principled ML methodology: generalisation and sample "70 "complexity, optimisation theory for learning, kernel and Bayesian methods, probabilistic models, causal "71 "inference, online learning, PAC learning, interpretability and fairness as formal problems. "72 "Architecture engineering is deep_learning_architectures."),73 topics=[("T12072", 1.0, "Machine Learning and Algorithms: PAC learning, generalisation bounds, VC dimension"),74 ("T11612", 0.8, "stochastic gradient optimisation, NTK, implicit regularisation"),75 ("T12814", 0.7, "Gaussian processes and Bayesian inference"),76 ("T11303", 0.6, "Bayesian modelling and causal inference"),77 ("T10845", 0.5, "advanced causal inference techniques"),78 ("T11901", 0.5, "Bayesian methods and mixture models"),79 ("T12535", 0.4, "ML and data classification (SVM, ensembles)"),80 ("T10637", 0.4, "clustering algorithms"),81 ("T12026", 0.4, "explainable AI"),82 ("T12536", 0.4, "topological and geometric data analysis"),83 ("T11652", 0.3, "imbalanced classification"),84 ("T11689", 0.3, "adversarial robustness (certified robustness)"),85 ("T10963", 0.3, "advanced optimisation algorithms"),86 ("T12761", 0.3, "data stream mining (concept drift)"),87 ("T12205", 0.3, "time series forecasting"),88 ("T13373", 0.2, "data mining and ML applications"),89 ("T12676", 0.2, "extreme learning machines"),90 ("T11512", 0.6, "anomaly detection techniques and applications"),91 ("T10820", 0.3, "fuzzy logic and control systems (soft computing)"),92 ("T11063", 0.3, "rough sets and fuzzy logic (granular computing, feature selection)")],93 subfields=[(1702, 0.3)],94 arxiv=[("stat.ML", 0.8), ("cs.LG", 0.4), ("math.ST", 0.3), ("math.OC", 0.2)],95 streams=[("conf/colt", 1.0), ("conf/alt", 1.0), ("journals/jmlr", 0.8), ("conf/aistats", 0.8), ("conf/uai", 0.8),96 ("journals/ml", 0.6), ("journals/tmlr", 0.4), ("conf/nips", 0.3), ("conf/icml", 0.3), ("conf/iclr", 0.2),97 ("journals/tnn", 0.3), ("journals/nn", 0.3), ("journals/pr", 0.2), ("conf/pkdd", 0.3), ("conf/kdd", 0.2),98 ("conf/icdm", 0.2), ("journals/pami", 0.2), ("journals/ijar", 0.4), ("journals/sac", 0.4),99 ("journals/csda", 0.3), ("journals/cstat", 0.2), ("journals/eswa", 0.1),100 ("journals/kbs", 0.1), ("journals/isci", 0.1), ("journals/ijon", 0.2), ("journals/nca", 0.1),101 ("journals/mlc", 0.3), ("journals/apin", 0.2), ("journals/air", 0.2), ("conf/ecai", 0.3), ("conf/ecai2", 0.3),102 ("conf/icmla", 0.5), ("conf/ssci", 0.4), ("conf/ictai", 0.3), ("conf/icic", 0.3), ("conf/aciids", 0.2),103 ("conf/kes", 0.2), ("journals/tcyb", 0.3), ("journals/tsmc", 0.2), ("journals/tfs", 0.3), ("journals/fss", 0.3),104 ("conf/fuzzIEEE", 0.3), ("journals/ijfs", 0.3), ("journals/jifs", 0.2), ("journals/ijis", 0.3),105 ("journals/ijcisys", 0.3), ("journals/comintsys", 0.3), ("journals/es", 0.3), ("journals/iasc", 0.2),106 ("journals/inffus", 0.3), ("journals/eaai", 0.2)],107 keywords=[108 (r"(?i)\b(generali[sz]ation (bounds?|error|gap|theory)|sample complexity|regret (bounds?|minimi[sz]ation|analysis)|minimax (rates?|optimal|lower)|PAC(-Bayes(ian)?)?( learning| bounds?)?|Rademacher|VC[- ]dimension|convergence (rates?|guarantees?|analysis)|statistical learning theory|computational learning theory|learnability|uniform convergence|excess risk)\b", 0.9, "both"),109 (r"(?i)\b(kernel (methods?|machines?|ridge|mean embedding)|reproducing kernel|RKHS|Gaussian process(es)?|variational inference|Bayesian (optimi[sz]ation|neural networks?|inference|deep learning|nonparametric)|probabilistic (models?|programming|graphical)|graphical models?|expectation[- ]maximi[sz]ation|Markov chain Monte Carlo|MCMC|Hamiltonian Monte Carlo|normali[sz]ing flows? for inference|conformal prediction|calibration)\b", 0.8, "both"),110 (r"(?i)\bcausal (inference|discovery|representation learning|effect estimation|structure learning|machine learning)\b|\bcounterfactual (reasoning|inference|fairness)\b|\bstructural causal models?\b|\btreatment effect\b", 0.8, "both"),111 (r"(?i)\b(online learning|bandits?|stochastic (gradient|optimi[sz]ation|approximation)|non-?convex optimi[sz]ation|convex optimi[sz]ation for learning|implicit (bias|regulari[sz]ation)|gradient (descent|flow) dynamics|saddle points?|loss landscapes?|mirror descent|optimal transport|Wasserstein)\b", 0.6, "both"),112 (r"(?i)\b(double descent|neural tangent kernel|NTK\b|lottery ticket|over-?parameteri[sz](ed|ation)|benign overfitting|grokking|feature learning (regime|theory)|mean-?field (limit|theory) of neural|infinite-?width|random features?)\b", 1.0, "both"),113 (r"(?i)\b(interpretab(le|ility)|explainab(le|ility)) (machine learning|AI|models?|methods?)\b|\bXAI\b|\bSHAP\b|\bLIME\b|\bfeature attribution\b|\bsaliency maps? for\b", 0.5, "both"),114 (r"(?i)\b(fairness (constraints?|metrics?|in machine learning)|algorithmic fairness|demographic parity|equali[sz]ed odds|fair (classification|representation))\b", 0.5, "both"),115 (r"(?i)\b(semi-?supervised|self-?supervised|unsupervised|active|transfer|meta|continual|lifelong|few-?shot|zero-?shot|multi-?task|weakly[- ]supervised|positive-?unlabeled|contrastive|representation|manifold|metric|dictionary|ensemble|boosting|bagging) learning\b", 0.4, "both"),116 (r"(?i)\b(distribution shift|out-?of-?distribution|domain generali[sz]ation|covariate shift|label noise|robust(ness)? (to|under|against) (noise|shift|corruption)|concept drift|anomaly detection|outlier detection|imbalanced (data|learning|classification)|dimensionality reduction|matrix (completion|factori[sz]ation)|tensor (decomposition|methods?)|sparse (coding|recovery)|compressed sensing)\b", 0.5, "both"),117 (r"(?i)\b(support vector machines?|SVMs?|random forests?|gradient boosting|XGBoost|LightGBM|k-?means|spectral clustering|Gaussian mixture|hidden Markov models?|HMMs?|logistic regression|decision trees?|nearest neighbou?rs?|kNN)\b", 0.5, "both"),118 (r"(?i)\bmachine learning\b|\bstatistical learning\b|\blearning theory\b", 0.4, "both"),119 ],120 exclusions=[],121 related=["deep_learning_architectures", "reinforcement_learning", "algorithms", "computational_complexity", "ai_safety", "foundation_models"],122 parent_of=[], child_of=[],123 provenance=("Hosts classical ML and its theory: OpenAlex splits them across T12072 (learning theory), T11612 (optimisation), "124 "T12814/T11303/T11901 (Bayesian), T12535/T10637 (classical methods). Venue weights follow venue purity "125 "(COLT/ALT/JMLR/AISTATS/UAI strong; NeurIPS/ICML partial)."),126),127dict(128 id="deep_learning_architectures", number=8, name="Deep Learning Architectures", area="ai_ml",129 definition=("Design, training and compression of neural network architectures as such: CNNs, RNNs, transformers and "130 "attention, graph neural networks, state-space models, normalisation and optimisers in practice, neural "131 "architecture search, pruning, quantisation and distillation, neuromorphic and spiking networks as models. "132 "Task-specific results go to the application category; hardware goes to ai_accelerators."),133 topics=[("T10320", 1.0, "Neural Networks and Applications: training, architectures, activation functions"),134 ("T11273", 1.0, "Advanced Graph Neural Networks"),135 ("T10036", 0.5, "Advanced Neural Network Applications: YOLO, lightweight networks, model compression"),136 ("T12676", 0.5, "extreme learning machines"),137 ("T12611", 0.4, "reservoir computing / echo state networks"),138 ("T11612", 0.3, "DNN training optimisation"),139 ("T11347", 0.3, "neural network stability and synchronisation"),140 ("T11307", 0.3, "PEFT and continual learning for deep nets"),141 ("T12205", 0.2, "time-series forecasting with deep nets"),142 ("T10502", 0.2, "neuromorphic computing (spiking models)"),143 ("T11206", 0.2, "model reduction and neural networks")],144 subfields=[(1702, 0.3)],145 arxiv=[("cs.NE", 1.0), ("cs.LG", 0.5), ("cs.CV", 0.1), ("cs.AI", 0.1)],146 streams=[("conf/iclr", 0.5), ("conf/nips", 0.4), ("conf/icml", 0.4), ("journals/tnn", 0.6), ("journals/nn", 0.6),147 ("conf/ijcnn", 0.6), ("journals/ijon", 0.5), ("journals/nca", 0.5), ("conf/icann", 0.6), ("conf/iconip", 0.5),148 ("journals/npl", 0.5), ("conf/esann", 0.5), ("journals/tmlr", 0.3), ("journals/pami", 0.2),149 ("journals/cin", 0.3), ("journals/tetci", 0.3), ("journals/tai", 0.3), ("journals/ficn", 0.2),150 ("journals/cogcom", 0.2), ("journals/aisy", 0.1), ("conf/aaai", 0.1)],151 keywords=[152 (r"(?i)\b(deep|convolutional|recurrent|graph|spiking|artificial|feed-?forward|residual|siamese|capsule|physics-?informed|liquid|binary|quanti[sz]ed|Hopfield|Kolmogorov-?Arnold) neural networks?\b", 0.8, "both"),153 (r"(?i)\b(CNNs?|RNNs?|LSTMs?|GRUs?|GNNs?|GCNs?|GATs?|SNNs?|MLPs?|DNNs?|ANNs?|KANs?|ResNets?|DenseNet|EfficientNet|MobileNet|U-?Net|ViTs?|Swin)\b", 0.7, "both"),154 (r"(?i)\btransformers?\b", 0.5, "both"),155 (r"(?i)\b(self-?|cross-?|multi-?head |sparse |linear |flash)?attention (mechanisms?|heads?|layers?|maps?|weights)\b|\battention is all\b", 0.7, "both"),156 (r"(?i)\bneural architecture search\b|\bNAS\b.*\b(architecture|search space|supernet)\b|\bdifferentiable architecture search\b|\bgraph neural networks?\b|\bnetwork architectures? (design|for deep learning)\b|\barchitectures? of (deep|convolutional|recurrent) neural\b|\bdeep (neural )?network (architectures?|design|compression)\b", 1.0, "both"),157 (r"(?i)\bAutoML\b|\bhyper-?parameter optimi[sz]ation\b", 0.6, "both"),158 (r"(?i)\b(model compression|knowledge distillation|network pruning|structured pruning|unstructured pruning|post-?training quanti[sz]ation|quanti[sz]ation-?aware training|low-?rank (factori[sz]ation|approximation) of (weights|layers)|weight sharing|parameter sharing|lightweight (networks?|architectures?|models?)|efficient (architectures?|networks?|transformers?))\b", 0.8, "both"),159 (r"(?i)\b(batch|layer|group|instance|weight|spectral) normali[sz]ation\b|\bdropout\b|\bresidual (connections?|learning|blocks?)\b|\bskip connections?\b|\bactivation functions?\b|\bReLU\b|\bGELU\b|\bSwiGLU\b|\bsoftmax\b", 0.7, "both"),160 (r"(?i)\b(mixture[- ]of[- ]experts|MoE\b|state[- ]space models?|SSMs?\b|Mamba|S4\b|RWKV|Hyena|linear (RNNs?|transformers?)|positional (encodings?|embeddings?)|RoPE|rotary|hypernetworks?|neural ODEs?|implicit (layers|neural representations?)|capsule networks?|equivariant (networks?|neural)|geometric deep learning|message passing (neural|networks?)|graph (convolution|attention|transformers?|pooling)|over-?smoothing)\b", 0.9, "both"),161 (r"(?i)\b(backpropagation|back-?prop|gradient (vanishing|exploding)|learning rate (schedul|warm-?up)|Adam(W)?\b|optimi[sz]ers? for (deep|neural)|weight (decay|initiali[sz]ation)|training (dynamics|instabilit)|loss functions? for|curriculum learning|data augmentation|mixup|CutMix|label smoothing|early stopping|hyper-?parameter (tuning|optimi[sz]ation))\b", 0.7, "both"),162 (r"(?i)\bdeep learning\b|\bneural networks?\b|\bneural nets?\b", 0.5, "both"),163 (r"(?i)\b(spiking|neuromorphic|reservoir computing|echo state|extreme learning machines?|ELM\b)\b", 0.5, "both"),164 ],165 exclusions=[r"(?is)^(?!.*(neural|deep learning|attention|language model|vision|encoder|decoder)).*\b(power|distribution|current|voltage|instrument|high-?voltage|dry-?type|oil-?immersed|pulse|flyback|isolation|phase-?shifting|step-?(up|down)) transformers?\b"],166 related=["ml_theory", "computer_vision", "nlp", "ai_accelerators", "foundation_models", "llm"],167 parent_of=[], child_of=[],168 provenance=("T10320 and T11273 are architecture-centred OpenAlex topics; cs.NE is the arXiv home of neural and "169 "evolutionary computation. 'Transformer' alone is weak and electrical senses are vetoed when no "170 "learning term co-occurs."),171),172dict(173 id="ai_safety", number=38, name="AI Safety and Alignment", area="ai_ml",174 definition=("Making AI systems safe, aligned, robust, fair, interpretable and governable: alignment and preference "175 "learning framed as a safety problem, red-teaming and jailbreaks, adversarial robustness and backdoors, "176 "hallucination and truthfulness, bias and fairness, mechanistic interpretability, evaluation of dangerous "177 "capabilities, AI governance and regulation. Generic system security is cybersecurity; XAI as a statistical "178 "method only is ml_theory."),179 topics=[("T10883", 1.0, "Ethics and Social Impacts of AI: AI safety, alignment, fairness, governance"),180 ("T11689", 0.6, "adversarial robustness, backdoors, prompt injection"),181 ("T12026", 0.5, "explainable AI / mechanistic interpretability"),182 ("T13851", 0.4, "law, AI and intellectual property"),183 ("T12262", 0.3, "hate speech and cyberbullying detection (harm mitigation)"),184 ("T11147", 0.3, "misinformation and its impacts"),185 ("T10525", 0.2, "human-automation interaction and safety"),186 ("T14307", 0.1, "bioethics and human rights")],187 subfields=[],188 arxiv=[("cs.CY", 0.5), ("cs.AI", 0.2), ("cs.LG", 0.1), ("cs.CL", 0.1), ("cs.CR", 0.1)],189 streams=[("conf/fat", 1.0), ("conf/aies", 1.0), ("journals/aiethics", 1.0), ("conf/satml", 1.0),190 ("journals/ais", 0.5), ("journals/natmi", 0.2), ("conf/nips", 0.1), ("conf/iclr", 0.1), ("conf/icml", 0.1),191 ("journals/tai", 0.2), ("journals/dai", 0.1), ("conf/sp", 0.1), ("conf/ccs", 0.1)],192 keywords=[193 (r"(?i)\bAI (safety|alignment|governance|regulation|ethics|risks?|act|auditing|accountability|harms?|misuse|incidents?|oversight|assurance|red[- ]teaming)\b", 1.0, "both"),194 (r"(?i)\b(aligned|alignment of|align(ing)?|misaligned|misalignment of) (large )?(language models?|LLMs?|AI (systems?|models?)|foundation models?|agents?)\b|\bsuper-?alignment\b|\bvalue alignment\b|\bhuman values?\b", 1.0, "both"),195 (r"(?i)\b(jail-?break(s|ing)?|red[- ]teaming|prompt injection|backdoor (attacks?|defenses?)|data poisoning|model (stealing|extraction|inversion)|membership inference|adversarial (prompts?|suffixes?)|refusal|safety (training|fine-?tuning|guardrails?|alignment|benchmarks?|evaluations?)|guardrails?|harmful (content|outputs?|responses?|requests?)|unsafe (outputs?|behavio(u)?rs?))\b", 1.0, "both"),196 (r"(?i)\badversarial (examples?|attacks?|robustness|training|perturbations?|defenses?|vulnerabilit)\b|\bcertified robustness\b|\brobustness (certification|verification)\b", 0.8, "both"),197 (r"(?i)\b(algorithmic|AI|model|machine learning|ML|LLM|gender|racial|demographic|social) (fairness|bias(es)?|discrimination|transparency|accountability|audits?|auditing)\b|\bbias (mitigation|amplification|in (LLMs?|language models?|AI))\b|\bdebiasing\b", 1.0, "both"),198 (r"(?i)\b(responsible|trustworthy|ethical|safe|harmless|honest|helpful|human-?centered|human-?compatible|beneficial) (AI|artificial intelligence|machine learning|LLMs?|agents?)\b", 1.0, "both"),199 (r"(?i)\b(mechanistic interpretability|sparse auto-?encoders? (for|on) (LLM|language|features)|activation (steering|patching)|circuit discovery|probing classifiers?|representation engineering|superposition hypothesis|induction heads|feature visuali[sz]ation|concept (bottleneck|activation vectors?)|model editing|knowledge editing)\b", 1.0, "both"),200 (r"(?i)\b(hallucinations?|truthfulness|factuality|sycophancy|deceptive alignment|deception|reward hacking|specification gaming|goal misgeneralization|scalable oversight|weak-?to-?strong|constitutional AI|RLHF|preference (learning|optimi[sz]ation)|human feedback|over-?optimi[sz]ation|situational awareness|sandbagging|power-?seeking|instrumental convergence|corrigib)\b", 0.8, "both"),201 (r"(?i)\b(existential risks?|catastrophic risks?|x-?risk|frontier (AI|models?)|dangerous capabilit(y|ies)|misuse of AI|dual-?use|AGI (safety|risk|alignment)|superintelligen|loss of control|AI (doom|pause)|compute governance|responsible scaling|model (cards?|evaluations? for safety)|capability evaluations?|system cards?)\b", 1.0, "both"),202 (r"(?i)\b(toxic(ity)?|hate speech|harmful|offensive|abusive) (language|content|speech|text|generation|detection)\b|\bcontent moderation\b|\b(misinformation|disinformation|fake news) (detection|generation|and LLMs?)\b", 0.5, "both"),203 (r"(?i)\b(watermark(ing)?|detect(ing|ion) of|detecting) (LLM|AI|machine)-?generated (text|content|images?)\b|\bAI-?text detection\b", 0.8, "both"),204 (r"(?i)\b(machine unlearning|right to be forgotten|differential privacy for (LLM|language)|memori[sz]ation (in|of) (LLM|language models?)|training data extraction)\b", 0.5, "both"),205 (r"(?i)\b(uncertainty (quantification|estimation|calibration)|out-?of-?distribution detection|distribution shift|safe (reinforcement learning|RL|exploration)|runtime monitoring of (neural|ML)|verification of neural networks?)\b", 0.4, "both"),206 (r"(?i)\b(ethics|ethical|societal (impact|implications)|social (impact|implications)|regulation|policy|governance|accountability|transparency|explainability|interpretability|fairness|bias|safety|risk)\b", 0.2, "both"),207 ],208 exclusions=[r"(?i)\b(multiple sequence|read|genome|sequence|image|spinal|wheel|dental|teeth|occlusal|fib(er|re)|beam|optical|antenna|shaft|laser) alignment\b(?!.*(AI|LLM|language model|RLHF|human values))"],209 related=["llm", "ml_theory", "cybersecurity", "privacy_preserving_computation", "ai_agents", "hci", "reinforcement_learning"],210 parent_of=[], child_of=[],211 provenance=("T10883 (field 33, Social Sciences) explicitly lists 'AI safety', 'AI alignment', 'algorithmic fairness'; "212 "T11689 and T12026 host the technical robustness/interpretability halves. FAccT, AIES, SaTML and AI & Ethics "213 "are the dedicated venues; the generic ethics/governance words are kept far below threshold."),214),215]216