"""Area ai_ml, part A: categories 1-4 (LLM, foundation models, multimodal AI, generative AI). Tuple conventions (converted to dicts by build_ontology.py): topics: (openalex_topic_id, weight, note) subfields:(subfield_id, weight) arxiv: (category, weight) streams: (dblp_stream_key, weight) keywords: (regex, weight, field) field in {"title", "abstract", "both"} """ CATEGORIES = [ dict( id="llm", number=1, name="Large Language Models", area="ai_ml", definition=("Research whose object is a transformer-based language model trained at scale: pre-training, " "instruction tuning, prompting and in-context learning, reasoning, retrieval augmentation, efficient " "inference, evaluation and benchmarks of LLMs. Excludes classical NLP tasks solved with small task-specific " "models (nlp) and non-language foundation models (foundation_models)."), topics=[("T10028", 1.0, "OpenAlex cluster hosting BERT / pre-trained LM / RAG / QA work: the de-facto LLM topic"), ("T11714", 0.4, "multimodal LLMs and vision-language models"), ("T12031", 0.3, "dialogue systems, LLM chatbots"), ("T10181", 0.3, "NLP techniques incl. neural MT and NLG"), ("T12128", 0.3, "chatbots in service interactions"), ("T14414", 0.2, "AI in education: mostly ChatGPT studies post-2023")], subfields=[(1702, 0.2)], arxiv=[("cs.CL", 0.5), ("cs.AI", 0.2), ("cs.LG", 0.2)], streams=[("conf/acl", 0.5), ("conf/emnlp", 0.5), ("conf/naacl", 0.5), ("conf/eacl", 0.4), ("conf/coling", 0.4), ("journals/tacl", 0.5), ("conf/iclr", 0.3), ("conf/nips", 0.3), ("conf/icml", 0.3), ("journals/tmlr", 0.3), ("conf/aaai", 0.2), ("conf/nlpcc", 0.3)], keywords=[ (r"(?i)\blarge language models?\b", 1.0, "both"), (r"(?i)\bLLMs?\b", 1.0, "both"), (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"), (r"(?i)\bpre-?trained language models?\b|\bPLMs?\b", 0.7, "both"), (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"), (r"(?i)\b(masked|autoregressive|causal|decoder-?only|encoder-?decoder) language model", 0.7, "both"), (r"(?i)\b(language model|transformer) (pre-?training|scaling|scaling laws?)\b", 0.7, "both"), (r"(?i)\b(parameter-?efficient fine-?tuning|PEFT\b|LoRA\b|adapters? for (language|LLM))\b", 0.6, "both"), (r"(?i)\bBERT\b|\bRoBERTa\b|\bT5\b|\bBART\b", 0.6, "both"), (r"(?i)\bhallucinat(ion|ions|e|es|ing)\b", 0.5, "both"), (r"(?i)\blanguage models?\b", 0.4, "both"), (r"(?i)\b(tokeni[sz](er|ation)|context (window|length)|KV[- ]cache|speculative decoding|mixture[- ]of[- ]experts)\b", 0.4, "both"), ], exclusions=[], related=["foundation_models", "nlp", "generative_ai", "ai_agents", "ai_safety"], parent_of=[], child_of=["foundation_models"], provenance=("Topics chosen by inspecting OpenAlex keyword lists (T10028 lists BERT, pre-trained language models, " "retrieval-augmented generation). Venues: *CL conferences and ML venues at partial weight because LLM " "work is spread across them. Keyword patterns reach the threshold alone, as required for the six " "AI categories OpenAlex cannot separate."), ), dict( id="foundation_models", number=2, name="Foundation Models", area="ai_ml", definition=("Large pre-trained models intended for adaptation to many downstream tasks in any modality: scaling " "laws, pre-training objectives, self-supervised representation learning at scale, transfer, fine-tuning " "and adapter methods, model releases and their evaluation. Language-specific work is llm (a child); " "vision-language work is multimodal_ai (a child)."), topics=[("T11307", 0.5, "domain adaptation, few-shot, parameter-efficient fine-tuning, continual learning"), ("T11714", 0.4, "vision-language models / CLIP"), ("T10028", 0.3, "pre-trained language models"), ("T10036", 0.2, "backbone networks, model compression"), ("T12422", 0.2, "foundation models in medical imaging / radiomics")], subfields=[(1702, 0.2)], arxiv=[("cs.LG", 0.4), ("cs.CV", 0.2), ("cs.CL", 0.2), ("cs.AI", 0.2)], streams=[("conf/nips", 0.3), ("conf/icml", 0.3), ("conf/iclr", 0.3), ("journals/tmlr", 0.3), ("conf/cvpr", 0.2), ("conf/iccv", 0.2), ("conf/eccv", 0.2), ("journals/natmi", 0.2)], keywords=[ (r"(?i)\bfoundation models?\b|\bfoundational models?\b", 1.0, "both"), (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"), (r"(?i)\bscaling laws?\b", 0.8, "both"), (r"(?i)\b(self-?supervised|contrastive|masked) (pre-?training|representation learning|learning at scale)\b", 0.7, "both"), (r"(?i)\b(pre-?train(ed|ing)) (on|at) (web|internet|large|massive|billion)[- ]scale", 0.7, "both"), (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"), (r"(?i)\b(general-?purpose|generalist|universal|all-?in-?one) (models?|agents?|representations?)\b", 0.6, "both"), (r"(?i)\b(transfer learning|fine-?tun(ing|ed?)|downstream tasks?|zero-?shot|few-?shot)\b", 0.4, "both"), (r"(?i)\b(adapters?|LoRA|parameter-?efficient|prompt tuning|emergent (abilities|capabilities))\b", 0.5, "both"), (r"(?i)\bworld models?\b", 0.5, "both"), ], exclusions=[], related=["llm", "multimodal_ai", "generative_ai", "deep_learning_architectures", "ml_theory"], parent_of=["llm", "multimodal_ai"], child_of=[], provenance=("No OpenAlex topic isolates foundation models; T11307 (PEFT, few-shot, continual) is the closest " "methodological host. Keyword patterns carry the category; the generic 'transfer learning/fine-tuning' " "terms are deliberately below threshold."), ), dict( id="multimodal_ai", number=3, name="Multimodal AI", area="ai_ml", definition=("Models and systems that jointly process two or more modalities (vision-language, audio-visual, " "video-text, 3D-language, sensors with language): vision-language models and multimodal LLMs, " "captioning, visual question answering, cross-modal retrieval and grounding. Unimodal vision is " "computer_vision; pure speech is speech_audio."), topics=[("T11714", 1.0, "Multimodal Machine Learning Applications: VLMs, MLLMs, VQA, captioning"), ("T10627", 0.4, "image and video retrieval incl. cross-modal"), ("T11439", 0.2, "video analysis and summarization"), ("T10824", 0.2, "image retrieval and classification"), ("T13155", 0.2, "digital communication and language (multimodal discourse)")], subfields=[(1707, 0.1)], arxiv=[("cs.MM", 0.6), ("cs.CV", 0.3), ("cs.CL", 0.3), ("cs.LG", 0.2)], streams=[("conf/mm", 0.6), ("journals/tmm", 0.6), ("conf/icmi", 0.6), ("journals/tomccap", 0.5), ("conf/icmcs", 0.5), ("conf/mir", 0.5), ("journals/mms", 0.5), ("journals/mta", 0.3), ("conf/cvpr", 0.2), ("conf/iccv", 0.2), ("conf/eccv", 0.2), ("conf/acl", 0.2), ("conf/emnlp", 0.2)], keywords=[ (r"(?i)\bmulti-?modal(ity)?\b", 0.7, "both"), (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"), (r"(?i)\b(VLMs?|MLLMs?|LVLMs?|LMMs?|VLAs?)\b", 1.0, "both"), (r"(?i)\bvisual question answering\b|\bVQA\b|\bvideo question answering\b", 1.0, "both"), (r"(?i)\b(image|video|dense|audio) captioning\b", 1.0, "both"), (r"(?i)\bcross-?modal(ity)?\b", 0.8, "both"), (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"), (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"), (r"(?i)\b(modalit(y|ies) (fusion|alignment|gap)|(early|late|intermediate) fusion of|sensor fusion with language)\b", 0.7, "both"), ], exclusions=[r"(?i)\bmultimodal (therapy|analgesi|treatment|intervention|pain|anesthesi|anaesthesi|imaging biomarkers?|prehabilitation|transport)\b"], related=["computer_vision", "nlp", "llm", "speech_audio", "generative_ai"], parent_of=[], child_of=["foundation_models"], provenance=("T11714 is the only OpenAlex topic centred on vision-language learning. Multimedia venues at partial " "weight (they also host unimodal media work). Medical 'multimodal' senses are vetoed."), ), dict( id="generative_ai", number=4, name="Generative AI", area="ai_ml", definition=("Models that synthesise new content (images, video, audio, 3D, text, code, molecules) and their training, " "sampling, evaluation, detection and misuse: GANs, VAEs, diffusion and flow models, autoregressive " "generators, text-to-X synthesis, deepfakes and AI-generated-content detection. LLM text generation as " "such is llm unless the contribution concerns the generative model or its outputs' authenticity."), topics=[("T10775", 1.0, "GANs, diffusion models, image/video generation"), ("T12357", 0.5, "digital media forensics: deepfake detection"), ("T11349", 0.2, "music technology (music generation)"), ("T10388", 0.2, "steganography / watermarking (of generated content)"), ("T14254", 0.2, "digital media and visual art (AI art)")], subfields=[], 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)], streams=[("conf/nips", 0.2), ("conf/icml", 0.2), ("conf/iclr", 0.2), ("conf/cvpr", 0.2), ("conf/iccv", 0.2), ("conf/eccv", 0.2), ("conf/siggraph", 0.2), ("conf/siggrapha", 0.2), ("journals/tog", 0.1)], keywords=[ (r"(?i)\bgenerative (AI|artificial intelligence|models?|modell?ing|adversarial networks?|pre-?trained|design|art)\b|\bGenAI\b|\bAIGC\b", 1.0, "both"), (r"(?i)\bGANs?\b|\b(conditional|cycle|style|Wasserstein)GAN\b|\bStyleGAN\w*\b", 0.9, "both"), (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"), (r"(?i)\bdiffusion models?\b", 0.8, "both"), (r"(?i)\btext-?to-?(image|video|3D|audio|speech|music|motion|SVG|scene)\b", 1.0, "both"), (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"), (r"(?i)\bdeep-?fakes?\b|\bface swap(ping)?\b|\bsynthetic media\b", 1.0, "both"), (r"(?i)\bvariational auto-?encoders?\b|\bVAEs?\b|\bVQ-?VAE\b|\bVQGAN\b", 0.7, "both"), (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"), (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"), (r"(?i)\bautoregressive (models?|generation|image|video)\b", 0.5, "both"), ], 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", r"(?is)^(?!.*(generative|GANs?\b|synthesis|generation|deepfake|VAE|AI-?generated)).*\bdiffusion (coefficients?|tensor|MRI|weighted|equations?|process(es)?|of innovations?|limited)\b"], related=["llm", "multimodal_ai", "computer_vision", "computer_graphics", "foundation_models", "speech_audio"], parent_of=[], child_of=[], provenance=("T10775 keywords explicitly list GANs, diffusion models, Stable Diffusion, text-to-image. Physics/finance " "senses of 'diffusion' are vetoed only when no generative-modelling term is present."), ), ]