Created yes · retained yes · inherited yes · invoked yes
Mechanism and system boundary must be read with the associated study; task gains alone do not prove better future improvement.
preprint · 2026-09-22
Recursive self-improvement of AI research agents
Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, Zhengyao Jiang
Why it matters here
Research-agent harness evolution with held-out evaluation and a separate outer-improver comparison.
AI-system improvementRecursive improvement evidence
What to keep in mind
- Company-authored evaluation; no independent replication recorded.
- September technical report differs from July blog; no numeric results merged across them.
Recursive-improvement study
Bounded structural L5; effective outer recursion unestablished.
| What changed | scaffold, prompt |
|---|---|
| What stayed fixed | Model weights; Outer selection rule; Task families and private scoring |
| Improved system used as optimizer later | yes |
| Generations attempted / accepted | 99 / 7 |
| Held-out transfer | External-task gains reported. |
| Resource accounting | Fixed per-evaluation constraints; eight-day main run. |
| Human contributions | Evaluator design, seed agents and resource limits. |
| Author claims | Harness research-efficiency gains and transfer. |
Recursive-improvement evidence
tracker evidence synthesisAssessment by RSI Tracker (Codex evidence synthesis) · 2026-09-26 · extraction review agent checked.
Budget matched: yes · independent evaluation: no
Improved harness transfer is reported; superior outer self-improvement is not established.
Not established by this particular source; no universal negative claim.
Resource budget:known · See study resource accounting; evaluation constraints are distinct from cumulative search cost.·Independence:known · Source-author report; no independent replication recorded.
Adaptivity
measuredSeven accepted rewrites; additional runs accept two and four.
Retention
not reportedNot recorded in this dataset.
Transfer
qualitativeFour held-out benchmarks.
Efficiency
qualitativePer-benchmark constraints; no decisive outer-loop efficiency gain.
Stability
qualitativeNoisy selection can admit harmful changes.
Meta-recursion
qualitativeEvolved AIDE47 invoked as outer improver; three seeds per arm are inconclusive.