| name | literature_review |
| description | Searches online for ML/AI papers relevant to a topic. Triggers on "find papers about X", "recent research on X", "any papers on this topic?". |
literature_review — paper search
Process
- Decompose the query. Break the user's topic into 2-4 core keyword combinations (e.g. "LLM inference efficiency" → "speculative decoding", "KV cache compression").
- Run online search. Use WebSearch targeting arXiv, Semantic Scholar, OpenReview, and conference proceedings (NeurIPS/ICML/ACL/CVPR etc.) first. No local PDF scanning — online only.
- Screen candidates. For each candidate paper, capture at minimum: publication year, venue (or arXiv-only status), authors/affiliation if available, one-line core contribution.
- Rank by relevance. Order by direct topical relevance → citation impact → recency. Don't bias purely toward newest — include at least one foundational paper for the area.
- Record output. Under
_papers/, create a search session note and record each paper as a Markdown link citation:
- [Authors et al., Year — Paper Title](arXiv/venue link): one-line contribution
Add any new terminology to _papers/glossary.md.
- Verify via source_verification. Don't finalize the output until each link is confirmed to actually point to the claimed paper.
Anti-Rationalization
- "The title looks relevant enough" — don't add to the list without at least checking the abstract.
- "Newest is always most important" — recency bias misses foundational prior work; always include a foundational paper.
- "It's an arXiv link so it's verified" — arXiv preprints may be un-peer-reviewed; state this explicitly.
Evidence
Search is complete when: every candidate has year/venue/contribution recorded, Markdown link citations actually resolve to the claimed paper, and at least one foundational paper is included.
Red Flags
- Adding to the list without reading the abstract
- Listing paper names with no link
- Presenting only a narrow time window (e.g. last 3 months) without disclosing that bias
