arxiv-research

# arXiv Research Skill Monitors arXiv for notable papers on LLMs, autonomous agents, and AI infrastructure. Scores papers by relevance, groups them by topic, and produces Markdown digests. - Covers `cs.AI`, `cs.CL`, `cs.LG`, `cs.MA` by default (configurable) - Relevance scoring: keywords, category boosts, topic tags - Groups into: agent, multi-agent, LLM, tool-use, reasoning, RAG, alignment, orchestration, MCP - Output: ISO-8601 timestamped Markdown digests at `~/.aibtc/arxiv-research/digests/` - No API key required — uses public arXiv Atom feed ## Usage ``` bun run arxiv-research/arxiv-research.ts <subcommand> [options] ``` ## Subcommands ### fetch Fetch recent papers from arXiv Atom API, score for LLM/agent relevance, and save to a local staging file. ```bash bun run arxiv-research/arxiv-research.ts fetch bun run arxiv-research/arxiv-research.ts fetch --categories "cs.CL,cs.MA" --max 100 ``` Options: - `--categories` — Comma-separated arXiv category codes. Default: `cs.AI,cs.CL,cs.LG,cs.MA` - `--max` — Maximum papers to fetch. Default: `50` Output: JSON summary with total/relevant counts and top 10 papers by relevance score. ### compile Compile a digest from the most recently fetched papers. Filters for relevance score ≥ 3, groups by topic, writes a Markdown file. ```bash bun run arxiv-research/arxiv-research.ts compile bun run arxiv-research/arxiv-research.ts compile --date 2026-03-19 ``` Options: - `--date` — Date label for the digest header. Default: today's date. Output: Writes `~/.aibtc/arxiv-research/digests/{ISO8601}_arxiv_digest.md` and prints summary JSON. ### list Show recent compiled digests. ```bash bun run arxiv-research/arxiv-research.ts list bun run arxiv-research/arxiv-research.ts list --limit 20 ``` Options: - `--limit` — Maximum entries to show. Default: `10` ## Relevance Scoring Papers are scored by matching title+abstract against topic signals: | Topic | Example Keywords | Weight | |-------|-----------------|--------| | agent | autonomous agent, AI agent, agent-based | 3–4 | | multi-agent | multi-agent | 4 | | LLM | large language model, LLM, GPT | 2–3 | | tool-use | tool use, function call | 3 | | reasoning | chain-of-thought, reasoning, planning | 2 | | RAG | retrieval-augmented, RAG | 2 | | alignment | RLHF, alignment | 2 | | orchestration | orchestrat* | 3 | | MCP | MCP, model context protocol | 3 | Category boosts: `cs.MA` +3, `cs.CL` +1, `cs.AI` +1. Minimum score for digest inclusion: 3. ## Output Format Digests are Markdown with: - Header: date, paper counts, categories - `## Highlights` — top 5 papers with scores - Per-topic sections with title, authors, arXiv link, abstract excerpt - `## Stats` table at the bottom