Your daily ArXiv firehose, filtered by GPT and actual taste

What it does

This is a daily ArXiv scanner that ingests RSS feeds, checks author lists against your watchlist, then runs the survivors through GPT-4 with a custom prompt describing your research interests. Papers get scored 1-10 on relevance and novelty; anything below your thresholds gets quietly discarded. Results land in Slack or a static GitHub Pages site.

The interesting bit

The whole thing runs on GitHub Actions for pennies—about seven cents a day to scan all of cs.CL as of February 2024. The clever cost-control is a two-stage filter: first an h-index cutoff culls obvious noise, then GPT-4 only sees the remainder. The prompt engineering is deliberately fussy, asking the model to cite specific criteria by number and distrust author claims of novelty.

Key highlights

• Author matching uses Semantic Scholar IDs, not just name strings, which helps with collisions and name changes
• JSONL output with per-paper comments explaining which criteria matched and why
• Designed to avoid re-announcing updated papers; only genuinely new submissions make it through
• Includes a debug mode to build regression tests for the filter when it makes mistakes
• Can run fully offline on a cheap VM with a single cron job if GitHub Actions feels too fickle

Caveats

• GPT-3.5 is explicitly labeled as debug-only; the README warns it “does not work well for this purpose”
• Without a Semantic Scholar API key, author lookups crawl; the README notes this but can’t fix it
• GitHub Actions workflows go inactive after 60 days on public repos, so the README nudges you toward making it private or self-hosting

Verdict

Worth it if you have specific, well-defined research interests and already know which authors to stalk. Skip it if you want serendipity or can’t articulate what you care about in a numbered list with counter-examples.