literature-review

Literature review

Before searching

• State the exact question the survey answers: task, setting, constraints ("parameter-efficient finetuning of sub-1B LLMs", not "efficient finetuning").
• State the recency window and why (e.g. "2023 onward; earlier work covered by survey X").
• State the inclusion criteria, so omissions are auditable.

Search procedure

1. Query multiple sources, not one: web search where available, plus direct API queries —
   • arXiv: curl -sL "https://export.arxiv.org/api/query?search_query=all:%22<terms>%22&max_results=20"
   • Semantic Scholar: curl -s "https://api.semanticscholar.org/graph/v1/paper/search?query=<terms>&fields=title,year,venue,abstract,citationCount,externalIds" — works without a key but rate-limits (HTTP 429); back off and retry, don't drop the source silently.
2. Use several distinct query formulations; different communities name the same idea differently.
3. Snowball: for the 2–3 most relevant papers found, follow their references and their citing papers (Semantic Scholar /paper/<id>/references and /citations).
4. Record every query run; the survey reports them so coverage can be audited.

Citation rules (non-negotiable)

• Never cite from memory. A paper appears in the survey only after its record (arXiv page, Semantic Scholar entry, or venue page) was fetched in this session. A paper that is remembered but cannot be found is reported as "recalled but unverified", not cited.
• Report what each paper shows, not what it claims: the dataset, metric, and number from its experiments, precise enough to be checked.
• Mark each entry as peer-reviewed (with venue) or preprint.
• Do not rank results that are not comparable (different datasets, metrics, or compute); say why they aren't.

Output

Save to papers/related-work/<topic>.md:

1. Question, scope, recency window, inclusion criteria.
2. Comparison table: method | venue+year (or preprint) | setting | metric | result | code available.
3. Gaps and open problems: what no surveyed paper addresses, stated specifically enough to seed a research question (input to the research-ideation skill).
4. Coverage limits: queries run, sources searched, where the survey may be incomplete.