| name | scout-rsi-papers |
| description | Sweep arXiv, OpenReview, and official lab blogs for new resources on recursive self-improvement in AI, dedupe against README.md, and produce candidate entries in the exact CONTRIBUTING.md format with section placement rationale. |
Scout RSI Papers
You are scouting new resources for the awesome-recursive-self-improvement list. The output of this skill is a set of candidate entries for maintainer review — never edit README.md directly.
Inputs
- Optional: a theme or README section to focus on (e.g. "test-time adaptation", "safety"). Default: a full-field sweep across all themes below.
- Optional: a time window (e.g. "last 3 months"). Default: since the most recent entries already in the list.
Sources (sweep in this order)
- arXiv — listings and search across
cs.AI, cs.LG, cs.CL (use https://arxiv.org/list/cs.AI/recent or arXiv search/API queries).
- OpenReview — accepted papers for the current ICLR/NeurIPS/ICML cycles, including the ICLR Workshop on Recursive Self-Improvement.
- Official lab blogs — Anthropic, Google DeepMind, OpenAI, Meta AI (FAIR), Sakana AI. Blog posts qualify only when they are the primary source for a system (e.g. AlphaEvolve).
- ACL Anthology — for published versions of NLP papers.
Query list by theme
Run searches combining the theme terms with year qualifiers for recency:
| Theme (target section) | Search strings |
|---|
| Self-refinement and reflection | "self-refine LLM", "self-correction language model", "intrinsic self-verification" |
| Test-time adaptation | "test-time training LLM", "test-time reinforcement learning", "inference-time self-improvement" |
| Experience learning and memory | "agent experience replay", "procedural memory LLM agent", "lifelong agent memory" |
| Self-generated data and self-rewarding | "self-rewarding language model", "self-play LLM training", "zero data reasoning", "synthetic data self-improvement" |
| Recursive decomposition and curricula | "automated curriculum LLM", "recursive problem decomposition", "self-generated curriculum" |
| Code evolution and algorithm discovery | "self-improving coding agent", "evolutionary code generation LLM", "agent modifies own code" |
| Self-evolving agents | "self-evolving agent", "self-referential agent", "agent skill library learning" |
| Multimodal and embodied | "embodied agent self-improvement", "multimodal self-rewarding" |
| Evaluation and benchmarks | "benchmark self-improving agents", "ML engineering agent benchmark", "AI research agent evaluation" |
| Safety, governance, limits | "reward hacking evaluation", "sabotage evaluation LLM", "model collapse recursive training", "self-modification safety" |
Process
- Sweep each source with the relevant queries (or only the focused theme if one was given).
- Screen each hit against the inclusion criteria in
CONTRIBUTING.md:
- Published or substantially updated 2022 onwards.
- Directly about a self-improvement loop (feedback, self-evaluation, self-generated data, code modification, memory, test-time adaptation, or governed updates) — not a generic agent framework.
- Primary source available (arXiv, ACL Anthology, OpenReview, official project page, official lab blog, official GitHub repository).
- Dedupe against the current list. For each candidate, check
README.md for:
- the exact URL,
- the arXiv ID in any form (
abs/, pdf/, versioned),
- the paper or system title (search a distinctive word from it).
- Canonicalise the URL: prefer arXiv
abs/ pages over PDFs, prefer published proceedings over preprints when both exist, strip tracking parameters, never use shorteners.
- Draft a one-sentence description that is neutral, factual, specific, and matches the linked source. Use UK spelling. Start with a verb where natural (matching list style: "Introduces…", "Surveys…", "Evaluates…").
- Place each candidate in exactly one README section, with a one-line rationale.
Output format
## Candidate entries (require maintainer review)
### <Section name>
- [Name](URL) (Venue Year) - One-sentence description.
- Rationale: <why it is in scope and why this section>
- Source checked: <arXiv / OpenReview / lab blog>, <year>
End with a short list of near-misses you excluded and the exclusion reason (helps the maintainer trust the sweep).
Guardrails
- Never edit
README.md — output candidates only.
- Never include secondary coverage (news articles, newsletters, social posts, summary blogs).
- Never inflate descriptions beyond what the source claims.
- If unsure whether a resource is in scope, list it under near-misses with your doubt stated, rather than as a candidate.
