preprint · 2026-09-21
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses (v3)
On this page
Why it matters here
This separate v3 snapshot reports task-level paired-bootstrap uncertainty and a second author-run repetition of the harness search in Appendix E.2. Both reported search runs improve every evaluated split over the unevolved harness. The additional repeatability evidence is useful, but unresolved GDPval accounting and conflicting abstract/result maxima remain visible; the v1 evidence is retained separately.
AI-system improvementRecursive improvement evidenceEvaluation integrity
What to keep in mind
- Distinct author-reported v3 snapshot; the already published v1 record remains unchanged.
- Appendix E.2 task-level paired bootstrap intervals describe finite-task evaluation uncertainty; the authors' second search run addresses only limited search repeatability. Neither is independent external replication.
- Appendix B.5 retains the unresolved GDPval accounting mismatch: 185 tasks, both presentation orders, but 204 comparisons per judge.
- The abstract says up to 4.3 out-of-distribution points, whereas Tables 1 and 6 report JobBench +4.7; do not silently reconcile this.
- Backbone weights are frozen. Lower policy-token use than unregularized evolution excludes total search/research costs and remains above the starting harness.
- No tracker autonomy or recursive-acceleration assessment is added.