1"""Area ai_ml, part A: categories 1-4 (LLM, foundation models, multimodal AI, generative AI).23Tuple conventions (converted to dicts by build_ontology.py):4 topics: (openalex_topic_id, weight, note)5 subfields:(subfield_id, weight)6 arxiv: (category, weight)7 streams: (dblp_stream_key, weight)8 keywords: (regex, weight, field) field in {"title", "abstract", "both"}9"""1011CATEGORIES = [12dict(13 id="llm", number=1, name="Large Language Models", area="ai_ml",14 definition=("Research whose object is a transformer-based language model trained at scale: pre-training, "15 "instruction tuning, prompting and in-context learning, reasoning, retrieval augmentation, efficient "16 "inference, evaluation and benchmarks of LLMs. Excludes classical NLP tasks solved with small task-specific "17 "models (nlp) and non-language foundation models (foundation_models)."),18 topics=[("T10028", 1.0, "OpenAlex cluster hosting BERT / pre-trained LM / RAG / QA work: the de-facto LLM topic"),19 ("T11714", 0.4, "multimodal LLMs and vision-language models"),20 ("T12031", 0.3, "dialogue systems, LLM chatbots"),21 ("T10181", 0.3, "NLP techniques incl. neural MT and NLG"),22 ("T12128", 0.3, "chatbots in service interactions"),23 ("T14414", 0.2, "AI in education: mostly ChatGPT studies post-2023")],24 subfields=[(1702, 0.2)],25 arxiv=[("cs.CL", 0.5), ("cs.AI", 0.2), ("cs.LG", 0.2)],26 streams=[("conf/acl", 0.5), ("conf/emnlp", 0.5), ("conf/naacl", 0.5), ("conf/eacl", 0.4), ("conf/coling", 0.4),27 ("journals/tacl", 0.5), ("conf/iclr", 0.3), ("conf/nips", 0.3), ("conf/icml", 0.3),28 ("journals/tmlr", 0.3), ("conf/aaai", 0.2), ("conf/nlpcc", 0.3)],29 keywords=[30 (r"(?i)\blarge language models?\b", 1.0, "both"),31 (r"(?i)\bLLMs?\b", 1.0, "both"),32 (r"(?i)\b(ChatGPT|GPT-?3(\.5)?|GPT-?4o?|GPT-?4\.\d|GPT-?5|LLaMA|Llama-? ?[234]|PaLM|InstructGPT|Qwen|Vicuna|Mixtral|DeepSeek-?(V\d|R1)|Claude ?[34]|Gemini ?(Pro|Ultra|1\.5|2))\b", 1.0, "both"),33 (r"(?i)\bpre-?trained language models?\b|\bPLMs?\b", 0.7, "both"),34 (r"(?i)\b(instruction[- ]tuning|instruction-?tuned|in-?context learning|chain[- ]of[- ]thought|retrieval[- ]augmented generation|RAG\b|prompt (engineering|tuning|learning)|prompting|RLHF|supervised fine-?tuning|SFT\b)\b", 0.8, "both"),35 (r"(?i)\b(masked|autoregressive|causal|decoder-?only|encoder-?decoder) language model", 0.7, "both"),36 (r"(?i)\b(language model|transformer) (pre-?training|scaling|scaling laws?)\b", 0.7, "both"),37 (r"(?i)\b(parameter-?efficient fine-?tuning|PEFT\b|LoRA\b|adapters? for (language|LLM))\b", 0.6, "both"),38 (r"(?i)\bBERT\b|\bRoBERTa\b|\bT5\b|\bBART\b", 0.6, "both"),39 (r"(?i)\bhallucinat(ion|ions|e|es|ing)\b", 0.5, "both"),40 (r"(?i)\blanguage models?\b", 0.4, "both"),41 (r"(?i)\b(tokeni[sz](er|ation)|context (window|length)|KV[- ]cache|speculative decoding|mixture[- ]of[- ]experts)\b", 0.4, "both"),42 ],43 exclusions=[],44 related=["foundation_models", "nlp", "generative_ai", "ai_agents", "ai_safety"],45 parent_of=[], child_of=["foundation_models"],46 provenance=("Topics chosen by inspecting OpenAlex keyword lists (T10028 lists BERT, pre-trained language models, "47 "retrieval-augmented generation). Venues: *CL conferences and ML venues at partial weight because LLM "48 "work is spread across them. Keyword patterns reach the threshold alone, as required for the six "49 "AI categories OpenAlex cannot separate."),50),51dict(52 id="foundation_models", number=2, name="Foundation Models", area="ai_ml",53 definition=("Large pre-trained models intended for adaptation to many downstream tasks in any modality: scaling "54 "laws, pre-training objectives, self-supervised representation learning at scale, transfer, fine-tuning "55 "and adapter methods, model releases and their evaluation. Language-specific work is llm (a child); "56 "vision-language work is multimodal_ai (a child)."),57 topics=[("T11307", 0.5, "domain adaptation, few-shot, parameter-efficient fine-tuning, continual learning"),58 ("T11714", 0.4, "vision-language models / CLIP"),59 ("T10028", 0.3, "pre-trained language models"),60 ("T10036", 0.2, "backbone networks, model compression"),61 ("T12422", 0.2, "foundation models in medical imaging / radiomics")],62 subfields=[(1702, 0.2)],63 arxiv=[("cs.LG", 0.4), ("cs.CV", 0.2), ("cs.CL", 0.2), ("cs.AI", 0.2)],64 streams=[("conf/nips", 0.3), ("conf/icml", 0.3), ("conf/iclr", 0.3), ("journals/tmlr", 0.3), ("conf/cvpr", 0.2),65 ("conf/iccv", 0.2), ("conf/eccv", 0.2), ("journals/natmi", 0.2)],66 keywords=[67 (r"(?i)\bfoundation models?\b|\bfoundational models?\b", 1.0, "both"),68 (r"(?i)\b(vision|visual|time[- ]series|protein|genomic|geospatial|medical|clinical|weather|climate|chemistry|molecular|tabular|graph|speech|audio|robotics?|world) foundation models?\b", 1.0, "both"),69 (r"(?i)\bscaling laws?\b", 0.8, "both"),70 (r"(?i)\b(self-?supervised|contrastive|masked) (pre-?training|representation learning|learning at scale)\b", 0.7, "both"),71 (r"(?i)\b(pre-?train(ed|ing)) (on|at) (web|internet|large|massive|billion)[- ]scale", 0.7, "both"),72 (r"(?i)\b(vision transformers?|ViT-?[BLHG]?|DINO(v2)?|segment anything|SAM ?2|masked auto-?encoders?|MAE\b|CLIP\b|SigLIP)\b", 0.6, "both"),73 (r"(?i)\b(general-?purpose|generalist|universal|all-?in-?one) (models?|agents?|representations?)\b", 0.6, "both"),74 (r"(?i)\b(transfer learning|fine-?tun(ing|ed?)|downstream tasks?|zero-?shot|few-?shot)\b", 0.4, "both"),75 (r"(?i)\b(adapters?|LoRA|parameter-?efficient|prompt tuning|emergent (abilities|capabilities))\b", 0.5, "both"),76 (r"(?i)\bworld models?\b", 0.5, "both"),77 ],78 exclusions=[],79 related=["llm", "multimodal_ai", "generative_ai", "deep_learning_architectures", "ml_theory"],80 parent_of=["llm", "multimodal_ai"], child_of=[],81 provenance=("No OpenAlex topic isolates foundation models; T11307 (PEFT, few-shot, continual) is the closest "82 "methodological host. Keyword patterns carry the category; the generic 'transfer learning/fine-tuning' "83 "terms are deliberately below threshold."),84),85dict(86 id="multimodal_ai", number=3, name="Multimodal AI", area="ai_ml",87 definition=("Models and systems that jointly process two or more modalities (vision-language, audio-visual, "88 "video-text, 3D-language, sensors with language): vision-language models and multimodal LLMs, "89 "captioning, visual question answering, cross-modal retrieval and grounding. Unimodal vision is "90 "computer_vision; pure speech is speech_audio."),91 topics=[("T11714", 1.0, "Multimodal Machine Learning Applications: VLMs, MLLMs, VQA, captioning"),92 ("T10627", 0.4, "image and video retrieval incl. cross-modal"),93 ("T11439", 0.2, "video analysis and summarization"),94 ("T10824", 0.2, "image retrieval and classification"),95 ("T13155", 0.2, "digital communication and language (multimodal discourse)")],96 subfields=[(1707, 0.1)],97 arxiv=[("cs.MM", 0.6), ("cs.CV", 0.3), ("cs.CL", 0.3), ("cs.LG", 0.2)],98 streams=[("conf/mm", 0.6), ("journals/tmm", 0.6), ("conf/icmi", 0.6), ("journals/tomccap", 0.5),99 ("conf/icmcs", 0.5), ("conf/mir", 0.5), ("journals/mms", 0.5), ("journals/mta", 0.3), ("conf/mmm", 0.5),100 ("journals/inffus", 0.2),101 ("conf/cvpr", 0.2), ("conf/iccv", 0.2), ("conf/eccv", 0.2), ("conf/acl", 0.2), ("conf/emnlp", 0.2)],102 keywords=[103 (r"(?i)\bmulti-?modal(ity)?\b", 0.7, "both"),104 (r"(?i)\b(vision[- ]language|vision[- ]and[- ]language|visual[- ]language|language[- ]vision|image[- ]text|video[- ]language|audio[- ]language|speech[- ]language|3D[- ]language|text[- ]to[- ]3D)[- ]?(models?|pre-?training|tasks?|alignment|understanding|learning|reasoning|navigation|action|retrieval|generation)?\b", 1.0, "both"),105 (r"(?i)\b(VLMs?|MLLMs?|LVLMs?|LMMs?|VLAs?)\b", 1.0, "both"),106 (r"(?i)\bvisual question answering\b|\bVQA\b|\bvideo question answering\b", 1.0, "both"),107 (r"(?i)\b(image|video|dense|audio) captioning\b", 1.0, "both"),108 (r"(?i)\bcross-?modal(ity)?\b", 0.8, "both"),109 (r"(?i)\b(audio-?visual|visual grounding|referring expression|visual dialog|visual reasoning|visual instruction tuning|LLaVA|BLIP-?2?|Flamingo|GPT-?4V|CLIP|ALIGN\b|CoCa)\b", 0.8, "both"),110 (r"(?i)\b(text-to-image|text-to-video|image-to-text|text-image|image-text) (retrieval|alignment|matching|generation|models?)\b", 0.8, "both"),111 (r"(?i)\b(modalit(y|ies) (fusion|alignment|gap)|(early|late|intermediate) fusion of|sensor fusion with language)\b", 0.7, "both"),112 ],113 exclusions=[r"(?i)\bmultimodal (therapy|analgesi|treatment|intervention|pain|anesthesi|anaesthesi|imaging biomarkers?|prehabilitation|transport)\b"],114 related=["computer_vision", "nlp", "llm", "speech_audio", "generative_ai"],115 parent_of=[], child_of=["foundation_models"],116 provenance=("T11714 is the only OpenAlex topic centred on vision-language learning. Multimedia venues at partial "117 "weight (they also host unimodal media work). Medical 'multimodal' senses are vetoed."),118),119dict(120 id="generative_ai", number=4, name="Generative AI", area="ai_ml",121 definition=("Models that synthesise new content (images, video, audio, 3D, text, code, molecules) and their training, "122 "sampling, evaluation, detection and misuse: GANs, VAEs, diffusion and flow models, autoregressive "123 "generators, text-to-X synthesis, deepfakes and AI-generated-content detection. LLM text generation as "124 "such is llm unless the contribution concerns the generative model or its outputs' authenticity."),125 topics=[("T10775", 1.0, "GANs, diffusion models, image/video generation"),126 ("T12357", 0.5, "digital media forensics: deepfake detection"),127 ("T11349", 0.2, "music technology (music generation)"),128 ("T10388", 0.2, "steganography / watermarking (of generated content)"),129 ("T14254", 0.2, "digital media and visual art (AI art)")],130 subfields=[],131 arxiv=[("cs.CV", 0.2), ("cs.LG", 0.3), ("cs.GR", 0.2), ("cs.SD", 0.1), ("cs.AI", 0.1), ("cs.MM", 0.1)],132 streams=[("conf/nips", 0.2), ("conf/icml", 0.2), ("conf/iclr", 0.2), ("conf/cvpr", 0.2), ("conf/iccv", 0.2),133 ("conf/eccv", 0.2), ("conf/siggraph", 0.2), ("conf/siggrapha", 0.2), ("journals/tog", 0.1)],134 keywords=[135 (r"(?i)\bgenerative (AI|artificial intelligence|models?|modell?ing|adversarial networks?|pre-?trained|design|art)\b|\bGenAI\b|\bAIGC\b", 1.0, "both"),136 (r"(?i)\bGANs?\b|\b(conditional|cycle|style|Wasserstein)GAN\b|\bStyleGAN\w*\b", 0.9, "both"),137 (r"(?i)\b(denoising diffusion|latent diffusion|stable diffusion|diffusion transformers?|DDPM|DDIM|score-?based generative|score matching|flow[- ]matching|rectified flow|normali[sz]ing flows?|consistency models?)\b", 1.0, "both"),138 (r"(?i)\bdiffusion models?\b", 0.8, "both"),139 (r"(?i)\btext-?to-?(image|video|3D|audio|speech|music|motion|SVG|scene)\b", 1.0, "both"),140 (r"(?i)\b(image|video|music|audio|speech|3D|motion|molecule|molecular|code|scene|face|avatar|talking[- ]head|layout|texture|graph) (generation|synthesis)\b", 0.7, "both"),141 (r"(?i)\bdeep-?fakes?\b|\bface swap(ping)?\b|\bsynthetic media\b", 1.0, "both"),142 (r"(?i)\bvariational auto-?encoders?\b|\bVAEs?\b|\bVQ-?VAE\b|\bVQGAN\b", 0.7, "both"),143 (r"(?i)\b(AI-?generated|machine-?generated|synthetic(ally)?[- ]generated|LLM-?generated) (content|images?|text|data|videos?|audio|art|code|faces)\b", 0.8, "both"),144 (r"(?i)\b(image (inpainting|outpainting|editing|to-image translation)|style transfer|super-?resolution|image restoration|controllable generation|ControlNet|DreamBooth|textual inversion|classifier-?free guidance|Midjourney|DALL-?E|Imagen|Sora)\b", 0.6, "both"),145 (r"(?i)\bautoregressive (models?|generation|image|video)\b", 0.5, "both"),146 ],147 exclusions=[r"(?is)^(?!.*(generative|GANs?\b|synthesis|generation|deepfake|VAE|AI-?generated)).*\b(drug|gas|molecular|thermal|anomalous|reaction-?|water|heat|ion|oxygen|particle|turbulent|surface|solute|vapou?r) diffusion\b",148 r"(?is)^(?!.*(generative|GANs?\b|synthesis|generation|deepfake|VAE|AI-?generated)).*\bdiffusion (coefficients?|tensor|MRI|weighted|equations?|process(es)?|of innovations?|limited)\b"],149 related=["llm", "multimodal_ai", "computer_vision", "computer_graphics", "foundation_models", "speech_audio"],150 parent_of=[], child_of=[],151 provenance=("T10775 keywords explicitly list GANs, diffusion models, Stable Diffusion, text-to-image. Physics/finance "152 "senses of 'diffusion' are vetoed only when no generative-modelling term is present."),153),154]155