LLM Research Survey 2026

Next.js (pages router) + React + TypeScript site for browsing 2026 NLP/ML conference papers, backed by SQLite (FTS5 search) and a keyword-scored topic taxonomy.

Run

npm install
npm run dev          # http://localhost:3000 (uses the bundled data/papers.db)

Rebuild the database

npm run data:fetch   # ICLR/ICML lists + ACL Anthology XML -> data/raw
npm run data:build   # python3 scripts/ingest.py -> data/papers.db

Sources: ICLR 2026 and ICML 2026 via papercopilot/paperlists (OpenReview); ACL 2026 (+Findings), EACL, LREC, CoNLL, SIGDIAL, TACL, CL via acl-org/acl-anthology. Not yet loaded: EMNLP, AACL, COLM, NeurIPS, AAAI 2026 (add a loader in scripts/ingest.py).

Types of work

A type of work = research area x role (what a paper contributes, and what it holds fixed vs varies). Vocabulary: scripts/worktypes.json.

• data/labels/o*.jsonl: per-paper labels for the 2,678 agent-related papers (types, fixes/varies, benchmarks, workflow, abstract-stated scores), produced by LLM reading of title+abstract.
• data/labels/type_notes.json: per-type synthesis (why members are one type, sub-variants, benchmark story, typical workflow).
• All other LLM papers get rule-based roles (marked rule).

npm run data:build       # base tables
npm run data:worktypes   # work-type tables (run on a local disk copy if your filesystem rejects SQLite locks)

Limits: labels use titles+abstracts only (no intro/related work/result tables); time order is by venue date; scores are only those stated in abstracts.

Pages

/ overview · /papers search + filters + taxonomy sidebar · /papers/[id] · /work-types (area x role matrix) · /work-types/[slug] (timeline, why-same, benchmarks, workflow, scores, papers) · /benchmarks · /taxonomy tree · /conferences/[slug]

Taxonomy

Defined in scripts/taxonomy.py (13 branches, ~48 leaves); edit patterns and re-run data:build. Heuristic: title x3, keywords x2, abstract x1. ML-venue papers with no language-model signal go to "Other ML".