Trending AI papers on arXiv (agent-curated)
Weekly velocity ranking of trending AI papers announced on arXiv. Four category queries against the official keyless arXiv query API (cat:cs.AI, cs.CL, cs.CV, cs.LG; announced in the trailing 7 complete days; 3 s between requests per the API Terms of Use) are merged and deduplicated on the base arXiv id, keeping the earliest announcement and the union of categories. Each paper is scored for interest — breadth-weighted recency: n_categories * exp(-age_days / 4) — and ranked fastest first, with deterministic first-match-wins AI-subtopic tags (agents, reasoning, evals, llm, finetuning, rl, quantization, multimodal, vision, audio, nlp, robotics, ml-theory, data, education, other), extractive two-sentence summaries of the abstract (LaTeX stripped, never generated prose), lead author + unioned affiliations, and an is_update flag (version > 1). Columns: ISO week, fetch timestamp (day-granular, UTC midnight), base arXiv id, version, is_update, title, extractive summary, AI subtopic, comma-joined categories, category count, authors, lead author, author count, pipe-joined distinct affiliations, announcement timestamp, days since announced, interest score, velocity rank, canonical arXiv URL, arXiv comment. Primary key: (week, arxiv_id). Cadence: weekly; each snapshot is the full trailing-7-day announcement universe, ranked freshest-and-broadest first. Nullability: comment, lead_affiliation and affiliations may be empty when the author supplied none; velocity_rank and interest_score are never null. Caveats: subtopic tags are keyword rules, not a classifier; affiliations are author-supplied strings; the interest score is an attention proxy, not a quality measure. Only descriptive metadata is stored (no full text), under the arXiv API Terms of Use (CC0 1.0 for metadata), so commercial_use = yes. Sample use: order by velocity_rank for the week's highest-interest AI papers, or filter ai_subtopic = 'agents'.
- machine-learning
- ai-research
- arxiv
- technology