Record · analyzed
Recursive self-improvement of AI research agents
An AI research agent tests modified versions of its own code on AI R&D tasks and keeps the best: seven successive improvements in an 8-day run.
1
Listed
Added to the list. How entries were checked when they were listed is in the journal.
2
Collected
| File | Source | SHA-256 |
|---|---|---|
| paper PDF, 2609.26457v12609.26457.pdf · 0.7 MB | arxiv.org/pdf/2609.26457v1 | 8ba79e0ad1028e43b0a47ee06162c9b3996ec90f7990b7ec35993444e505409b |
To check a file, download it from its source and compare its SHA-256 with the one above. Collected copies are kept in the editor's RSI system; manifest: design-life/literature/rsi-design/sources.json.
3
Analyzed
Quoted verbatim from design-life/literature/rsi-workflow-improvement/comparison.md (2026-10-07); design-life/research/rsi-design/evidence-review-2026-10-28.md (2026-10-28). Parts about applying the paper to the editor's private work are left out.
From “Evidence for improving reusable agent workflows” (2026-10-07)
Inspected edition: Recursive self-improvement of AI research agents, v1, 22 September 2026; PDF pp. 2–7, 10–14
Mechanism and reported evidence: AIDE² changes its research harness, with equal-budget private grading and four external benchmarks. An eight-day run generated 99 proposals and accepted seven improvements; two further runs accepted two and four. The strongest discovered agent matches or exceeds a human-engineered baseline across the four external benchmarks.
From “Evidence for effective RSI implementation” (2026-10-28)
Design choice: Let agents choose execution dynamically
Supporting evidence: AIDE² discovers useful search and context policies. AIDE²
Challenge or limit: Agentless provides a competitive fixed-procedure baseline; dynamic control needs to earn its extra cost. These are different studies, not a matched contest between their systems. Agentless
The evidence also leaves other productive changes available. DGM changes agent tools and software under fixed task criteria. AIDE² changes search and context policies. ACE changes reusable knowledge contents. AFlow changes prompts and connections. These interventions overlap, but they should not all be relabeled as improvements to the semantic task schema. DGM
4
Replicated
Not done yet. Not attempted yet.
5
Audited
Not done yet. Not attempted yet.
The editor's view
None recorded yet.