13924f8985
- Next.js/React/TypeScript app: overview, paper explorer with FTS5 search, paper detail, global taxonomy tree with per-conference tag breakdown - Python ingestion (ICLR/ICML 2026 via paperlists, ACL Anthology 2026) into data/papers.db, plus keyword-scored taxonomy (scripts/taxonomy.py) Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bhs9LU8eMixk6g1kb1d3j7
82 lines
13 KiB
Python
82 lines
13 KiB
Python
"""Taxonomy tree for LLM research (2026) + keyword scoring rules.
|
|
|
|
Each leaf: (id, label, [regex, ...]). Patterns are matched case-insensitively against
|
|
title (x3), author keywords (x2) and abstract (x1, capped).
|
|
"""
|
|
|
|
TREE = [
|
|
("arch", "Architectures & Foundation Models", [
|
|
("arch/attention", "Attention, SSMs & MoE", [r"\battention (mechanism|head|sink)s?\b", r"\bstate[- ]space\b", r"\bmamba\b", r"\blinear attention\b", r"\bmixture[- ]of[- ]experts\b", r"\bmoe\b", r"\bsparse attention\b", r"\btransformer architecture"]),
|
|
("arch/diffusion", "Diffusion & Non-autoregressive LMs", [r"\bdiffusion (language )?models?\b", r"\bmasked diffusion\b", r"\bdiscrete diffusion\b", r"\bdllms?\b", r"\bnon-?autoregressive\b", r"\bparallel decoding\b"]),
|
|
("arch/tokenization", "Tokenization & Embeddings", [r"\btokeni[sz]\w*", r"\bbyte[- ]level\b", r"\bsubword\b", r"\b(text|sentence|document) embeddings?\b", r"\bembedding models?\b", r"\bcontrastive (text )?embedding"]),
|
|
("arch/multimodal", "Multimodal & Vision-Language Models", [r"\bmultimodal\b", r"\bmulti-modal\b", r"\bvision[- ]language\b", r"\bvlms?\b", r"\bmllms?\b", r"\bvideo (understanding|language)\b", r"\bspeech (language )?models?\b", r"\baudio[- ]language\b", r"\bimage captioning\b", r"\bvisual question answering\b"]),
|
|
("arch/longctx", "Long Context & Memory", [r"\blong[- ]context\b", r"\bcontext (window|length)\b", r"\bcontext extension\b", r"\blong sequences?\b", r"\bneedle[- ]in[- ]a[- ]haystack\b", r"\blong[- ]term memory\b"]),
|
|
]),
|
|
("train", "Pretraining, Data & Adaptation", [
|
|
("train/pretrain", "Pretraining & Scaling Laws", [r"\bpre-?training\b", r"\bscaling laws?\b", r"\bcontinual pre-?training\b", r"\bmid-?training\b", r"\bcompute[- ]optimal\b", r"\blearning rate schedule"]),
|
|
("train/data", "Data Curation & Synthetic Data", [r"\bsynthetic data\b", r"\bdata (selection|curation|filtering|mixture|quality)\b", r"\bdata synthesis\b", r"\bdata augmentation\b", r"\bdata-?efficient\b", r"\bcorpus (construction|curation)"]),
|
|
("train/sft", "Instruction Tuning & SFT", [r"\binstruction[- ]tun\w*", r"\bsupervised fine-?tuning\b", r"\bsft\b", r"\binstruction[- ]following\b", r"\bfine-?tuning\b"]),
|
|
("train/peft", "Parameter-Efficient Fine-Tuning", [r"\blora\b", r"\bparameter-efficient\b", r"\bpeft\b", r"\badapters?\b", r"\blow-rank adapt\w*", r"\bprompt tuning\b", r"\bprefix[- ]tuning\b"]),
|
|
("train/continual", "Continual Learning & Model Merging", [r"\bcontinual learning\b", r"\bcatastrophic forgetting\b", r"\bmodel merging\b", r"\bweight (merging|averaging|interpolation)\b", r"\btask arithmetic\b", r"\blifelong\b"]),
|
|
("train/distill", "Distillation & Small Models", [r"\bknowledge distillation\b", r"\bdistill\w*", r"\bsmall language models?\b", r"\bslms?\b", r"\bon-?device\b"]),
|
|
]),
|
|
("align", "Alignment & Post-Training", [
|
|
("align/pref", "Preference Optimization & Reward Models", [r"\brlhf\b", r"\bdpo\b", r"\bpreference (optimization|learning|data|alignment)\b", r"\breward models?\b", r"\breward model(l)?ing\b", r"\bhuman feedback\b", r"\bsimpo\b", r"\bpreference pairs?\b"]),
|
|
("align/rlvr", "RL for Reasoning (RLVR / GRPO)", [r"\bgrpo\b", r"\brlvr\b", r"\bverifiable rewards?\b", r"\breinforcement learning (for|with) (llms?|reasoning|language)", r"\bppo\b", r"\bpolicy optimi[sz]ation\b", r"\bprocess reward\b", r"\boutcome reward\b", r"\brl (training|fine-?tuning)\b"]),
|
|
("align/values", "Values, Sycophancy & Pluralistic Alignment", [r"\bsycophan\w*", r"\bvalue alignment\b", r"\bpluralistic\b", r"\bhuman values\b", r"\bconstitutional\b", r"\bhelpful(ness)? and harmless\b", r"\bpersona\b", r"\bethical (reasoning|alignment)"]),
|
|
]),
|
|
("reason", "Reasoning & Code", [
|
|
("reason/cot", "Chain-of-Thought & Test-Time Compute", [r"\bchain[- ]of[- ]thought\b", r"\bcot\b", r"\btest-?time (compute|scaling)\b", r"\binference-?time (compute|scaling)\b", r"\blong cot\b", r"\breasoning models?\b", r"\bself-?consistency\b", r"\bthinking (tokens|budget|models?)\b", r"\boverthinking\b", r"\blatent reasoning\b", r"\btree[- ]of[- ]thought", r"\breasoning\b", r"\binference-?time\b"]),
|
|
("reason/math", "Math & Formal Reasoning", [r"\bmath(ematical|ematics)?\b", r"\btheorem prov\w*", r"\bformal (verification|proof|reasoning)\b", r"\blean\b", r"\bgsm8k\b", r"\bmath-?500\b", r"\baime\b", r"\bolympiad\b", r"\bautoformali[sz]ation\b"]),
|
|
("reason/code", "Code Generation & Software Engineering", [r"\bcode generation\b", r"\bcode (llms?|models?|completion|repair)\b", r"\bswe-?bench\b", r"\bprogram synthesis\b", r"\bsoftware engineering\b", r"\bcoding (agents?|benchmarks?|tasks?)\b", r"\bcompetitive programming\b", r"\bunit tests?\b", r"\bbug (fixing|localization)\b", r"\btext-to-sql\b"]),
|
|
("reason/planning", "Planning, Logic & Commonsense", [r"\bplanning\b", r"\blogical reasoning\b", r"\bcommonsense\b", r"\bcausal (reasoning|inference)\b", r"\bsymbolic\b", r"\bneuro-?symbolic\b", r"\bmulti-?hop\b", r"\bspatial reasoning\b", r"\btemporal reasoning\b", r"\babstract reasoning\b", r"\barc-?agi\b"]),
|
|
]),
|
|
("agents", "Agents & Tool Use", [
|
|
("agents/tools", "Tool Use & Function Calling", [r"\btool (use|learning|calling|retrieval|call\w*)\b", r"\bfunction[- ]calling\b", r"\bmodel context protocol\b", r"\bmcp\b", r"\bapi calls?\b", r"\btool-?augmented\b", r"\btool-?integrated\b"]),
|
|
("agents/autonomous", "Autonomous, Web & GUI Agents", [r"\bweb agents?\b", r"\bgui agents?\b", r"\bcomputer[- ]use\b", r"\bllm agents?\b", r"\bagentic\b", r"\bautonomous agents?\b", r"\blanguage agents?\b", r"\bbrowser agents?\b", r"\bmobile agents?\b", r"\bdeep research\b", r"\bagent (benchmark|trajector)\w*", r"\bagents?\b", r"\blong-horizon\b", r"\bmulti-turn (interaction|agent)"]),
|
|
("agents/multi", "Multi-Agent Systems", [r"\bmulti-?agent\b", r"\bagent collaboration\b", r"\bagent communication\b", r"\bdebate\b", r"\bsocial simulation\b", r"\bsociety of agents\b"]),
|
|
("agents/memory", "Agent Memory & Self-Improvement", [r"\bagent(ic)? memory\b", r"\bself-?(evolving|improv\w*|refine\w*|reflection)\b", r"\bexperience (replay|learning)\b", r"\bskill (library|learning)\b", r"\bmemory (management|system|retrieval)", r"\brefinement\b", r"\bself-?correct\w*"]),
|
|
]),
|
|
("knowledge", "Retrieval, Knowledge & Factuality", [
|
|
("knowledge/rag", "Retrieval-Augmented Generation", [r"\bretrieval[- ]augmented\b", r"\brag\b", r"\bretriev(al|er)s?\b", r"\bdense retrieval\b", r"\breranking\b", r"\bre-?rankers?\b", r"\bsearch (agents?|engines?)\b", r"\bgraphrag\b", r"\bopen-domain (qa|question)"]),
|
|
("knowledge/halluc", "Hallucination & Factuality", [r"\bhallucinat\w*", r"\bfactuality\b", r"\bfactual(ity)? (errors?|consistency|accuracy)\b", r"\bfaithfulness\b", r"\bfact[- ]check\w*", r"\bgrounded?ness\b", r"\battribution\b", r"\bcitation\b"]),
|
|
("knowledge/editing", "Knowledge Editing & Probing", [r"\bknowledge editing\b", r"\bmodel editing\b", r"\bknowledge (conflicts?|probing|storage|acquisition|injection)\b", r"\bparametric knowledge\b", r"\bknowledge graphs?\b", r"\bknowledge-intensive\b", r"\bfactual knowledge\b"]),
|
|
]),
|
|
("eval", "Evaluation & Benchmarks", [
|
|
("eval/bench", "Benchmarks & Datasets", [r"\bbenchmarks?\b", r"\bevaluat(e|ing|ion)\b", r"\bdatasets?\b", r"\bleaderboards?\b", r"\bevaluation suite\b", r"\btestbed\b", r"\bnew corpus\b", r"\bannotated corpus\b"]),
|
|
("eval/judge", "LLM-as-a-Judge & Evaluation Methodology", [r"\bllm[- ]as[- ]a[- ]judge\b", r"\bllm-?judges?\b", r"\bautomatic evaluation\b", r"\bevaluation metrics?\b", r"\bhuman evaluation\b", r"\bevaluator\b", r"\bjudge models?\b", r"\barena\b", r"\bpairwise comparison"]),
|
|
("eval/robust", "Robustness, Contamination & Calibration", [r"\bdata contamination\b", r"\bcontaminat\w*", r"\brobustness\b", r"\bcalibrat\w*", r"\buncertainty\b", r"\bconfidence estimation\b", r"\bprompt sensitivity\b", r"\bdistribution shift\b", r"\bout-of-distribution\b", r"\bgeneralization\b"]),
|
|
]),
|
|
("interp", "Interpretability & Analysis", [
|
|
("interp/mech", "Mechanistic Interpretability", [r"\bmechanistic\b", r"\binterpretab\w*", r"\bsparse autoencoders?\b", r"\bsaes?\b", r"\bcircuits?\b", r"\battention heads?\b", r"\blogit lens\b", r"\bactivation patching\b", r"\bneurons?\b", r"\bfeatures? (representation|visuali)", r"\bexplainab\w*"]),
|
|
("interp/steering", "Representation Analysis & Steering", [r"\bsteering vectors?\b", r"\bactivation (steering|engineering|addition)\b", r"\brepresentation engineering\b", r"\bprobing\b", r"\blinear representation\b", r"\blatent space\b", r"\bresidual stream\b", r"\bhidden states?\b"]),
|
|
("interp/icl", "In-Context Learning & Emergent Abilities", [r"\bin-?context learning\b", r"\bicl\b", r"\bfew-shot (prompting|learning)\b", r"\bemergent\b", r"\bprompt(ing)? (engineering|strateg|optimi[sz])\w*", r"\bdemonstrations?\b", r"\binduction heads?\b", r"\btheor(y|etical) (analysis|understanding) of (transformers?|llms?)"]),
|
|
]),
|
|
("efficiency", "Efficiency & Systems", [
|
|
("efficiency/compress", "Quantization, Pruning & Compression", [r"\bquanti[sz]\w*", r"\bpruning\b", r"\bsparsity\b", r"\bmodel compression\b", r"\bweight-?only\b", r"\blow-bit\b", r"\bkv cache (compression|quantization|eviction)"]),
|
|
("efficiency/inference", "Inference Acceleration & Serving", [r"\bspeculative decoding\b", r"\bkv[- ]cache\b", r"\binference (efficiency|acceleration|latency|speed)\b", r"\bserving\b", r"\bthroughput\b", r"\bdecoding (speed|strategy|algorithm)\b", r"\bearly exit\b", r"\bprefill\b", r"\btoken (pruning|merging|eviction)\b", r"\befficient (inference|decoding)\b"]),
|
|
("efficiency/training", "Efficient & Distributed Training", [r"\bdistributed training\b", r"\bmemory-?efficient\b", r"\bgpu\b", r"\btraining efficiency\b", r"\boptimizers?\b", r"\bmuon\b", r"\badam\w*", r"\bgradient (checkpointing|compression)\b", r"\bparallelism\b", r"\bfp8\b", r"\bmixed[- ]precision\b"]),
|
|
]),
|
|
("safety", "Safety, Security & Society", [
|
|
("safety/jailbreak", "Jailbreaks, Red-Teaming & Safety Alignment", [r"\bjailbreak\w*", r"\bred[- ]team\w*", r"\bsafety (alignment|training|guardrails?|evaluation|benchmark)\b", r"\bharmful\b", r"\brefusals?\b", r"\bguardrails?\b", r"\badversarial (attacks?|prompts?|suffix)\b", r"\bsafeguards?\b", r"\bmisuse\b"]),
|
|
("safety/security", "Security: Injection, Backdoors & Watermarking", [r"\bprompt injection\b", r"\bbackdoor\w*", r"\bdata poisoning\b", r"\bpoison\w*", r"\bwatermark\w*", r"\bmachine-?generated text detection\b", r"\bai-?generated text\b", r"\bmalicious\b", r"\bvulnerabilit\w*", r"\bmodel stealing\b"]),
|
|
("safety/privacy", "Privacy, Memorization & Unlearning", [r"\bprivacy\b", r"\bmemori[sz]ation\b", r"\bmembership inference\b", r"\bunlearning\b", r"\bdifferential(ly)? privat\w*", r"\bcopyright\b", r"\bdata extraction\b", r"\bpii\b", r"\bfederated\b"]),
|
|
("safety/fairness", "Bias, Fairness & Culture", [r"\bbias(es)?\b", r"\bfairness\b", r"\bstereotyp\w*", r"\bcultur\w*", r"\bsocial bias\b", r"\bdemographic\b", r"\bgender\b", r"\bpolitical (bias|leaning|orientation)\b", r"\btoxicity\b", r"\bhate speech\b", r"\bpersonali[sz]ation\b", r"\bdiversity\b"]),
|
|
]),
|
|
("multiling", "Multilingual & Low-Resource", [
|
|
("multiling/multi", "Multilingual LLMs & Cross-Lingual Transfer", [r"\bmultilingual\b", r"\bcross-?lingual\b", r"\bcode-?switch\w*", r"\blanguage (transfer|confusion)\b", r"\blanguage-specific\b", r"\bnon-english\b"]),
|
|
("multiling/lowres", "Low-Resource Languages", [r"\blow-?resource\b", r"\bunder-?resourced\b", r"\bindigenous\b", r"\bendangered\b", r"\bafrican languages?\b", r"\bindic\b", r"\bdialects?\b", r"\barabic\b", r"\bswahili\b", r"\bhindi\b", r"\bvietnamese\b", r"\bindonesian\b"]),
|
|
("multiling/mt", "Machine Translation", [r"\bmachine translation\b", r"\btranslation (quality|evaluation|models?)\b", r"\bnmt\b", r"\bsimultaneous translation\b", r"\bspeech translation\b", r"\bllm-?based translation\b"]),
|
|
]),
|
|
("apps", "Applications & Domains", [
|
|
("apps/science", "Biomedical, Clinical & Scientific", [r"\bbiomedical\b", r"\bclinical\b", r"\bmedical\b", r"\bhealth(care)?\b", r"\bscientific (literature|discovery|papers?|reasoning)\b", r"\bchemistry\b", r"\bprotein\b", r"\bdrug\b", r"\bpatients?\b", r"\bmaterials?\b", r"\bbiolog\w*", r"\bscience\b", r"\bradiology\b"]),
|
|
("apps/domain", "Legal, Finance, Education & Social/Cognitive Science", [r"\blegal\b", r"\bfinanc\w*", r"\beducation(al)?\b", r"\btutoring\b", r"\bstudents?\b", r"\bcomputational social science\b", r"\bpolitical\b", r"\blaw\b", r"\beconom\w*", r"\bpsycholog\w*", r"\bmental health\b", r"\bhuman-?ai (interaction|collaboration)\b", r"\btheory of mind\b", r"\bpersuasi\w*", r"\bpersonality\b", r"\bhuman rights\b", r"\bcognitive\b", r"\bsocial\b"]),
|
|
("apps/dialogue", "Dialogue, Role-Play & Generation", [r"\bdialogue\b", r"\bconversation\w*", r"\bchatbots?\b", r"\brole-?play\w*", r"\bstory\b", r"\bcreative writing\b", r"\btext generation\b", r"\bsummari[sz]ation\b", r"\bdata-to-text\b", r"\buser simulat\w*", r"\bempath\w*", r"\bemotion\w*"]),
|
|
("apps/classic", "Classical NLP Tasks", [r"\bnamed entity\b", r"\bner\b", r"\bparsing\b", r"\bsentiment\b", r"\brelation extraction\b", r"\binformation extraction\b", r"\bevent extraction\b", r"\bsemantic role\b", r"\bcoreference\b", r"\btext classification\b", r"\bquestion answering\b", r"\bnatural language inference\b", r"\bpos tagging\b", r"\bmorpholog\w*", r"\bsyntax\b", r"\bsyntactic\b", r"\bsemantic\w*", r"\blinguistic\w*", r"\bpragmatic\w*", r"\bdiscourse\b", r"\bspeech recognition\b", r"\basr\b"]),
|
|
]),
|
|
("other", "Other", [
|
|
("other/ml", "Other ML (non-language)", []),
|
|
("other/unclassified", "Unclassified", []),
|
|
]),
|
|
]
|