RRSI: Regularizing Recursive Self-Improvement Boosts OOD Agent Benchmarks Up to 4.7 Points
burny_tech · x · 2026-09-28
Researchers from Google Cloud AI Research, UNC, Stanford and WUSTL introduce RRSI (Regularized Recursive Self-Improvement of Agent Harnesses), addressing overfitting in recursively self-improving agents.
Problem: An LLM agent's capability is largely magnified by its harness (prompts, control flow, tooling, memory, context management). Existing methods iteratively edit harness components for recursive self-improvement, but often memorize training tasks—in-distribution gains shrink or vanish on out-of-distribution benchmarks.
Method: Regularization principles constrain both proposal and selection:
- A proposer with a temporally annealed edit budget limits bundled changes per candidate and encourages unexplored trajectories;
- A selector with a critic screens benchmark-specific proposals, while a pruner removes changes that are too small, too expensive, or no longer useful.
Results: Across 8 benchmarks spanning coding, agentic workspace, and engineering design, RRSI gains up to 14.1 points on its evolved split and up to 4.7 points on 5 OOD benchmarks, while using 30% fewer policy tokens than unregularized evolution.
More from coding & agent
- Liquid AI makes the case for edge agents: 350M-2.6B models with no per-token bill — JosephJacks_ · 2026-09-28
- X's invite-only bot creator program raises calls for agent usage dashboards — RachelVT42 · 2026-09-28
- A prompt trick to make quota-limited coding agents resume after interruptions — umeshonai · 2026-09-28
- Local two-model writing pipeline: Qwen outlines, Gemma polishes, zero edits — dushyant30suthar · 2026-09-28
- Dev runs Bonsai 2 locally on Mac Studio to diff medication lists in 5.75 seconds — MaziyarPanahi · 2026-09-28
- Fixing Claude Opus 5.5 long-context drift with an external SESSION_STATE.md protocol — dr_cintas · 2026-09-28