RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Google Research's RRSI regularizes recursive evolution of LLM agent harnesses, gaining up to 14.1 in-distribution and 4.7 out-of-distribution points while using 30% fewer policy tokens.

Google Research introduces Regularized Recursive Self-Improvement (RRSI), a method that constrains how LLM agent harnesses evolve their prompts, control flow, tools, memory, and context so recursive self-improvement does not overfit training tasks. A temporally annealed edit budget and history-based exploration that favors unexplored evolution paths limit candidate changes, while a critic and a pruner reject benchmark-specific, tiny, costly, or obsolete edits. Across eight coding, workspace, and engineering-design benchmarks, both the Hugging Face daily-paper entry (2026-09-20) and the arXiv report (2026-09-21) state that RRSI gains up to 14.1 points in-distribution and 4.7 points on five out-of-distribution test sets, using 30% fewer policy tokens than unregularized evolution. The Hugging Face report adds that code is released under Google Research. The two reports agree on attribution and all headline figures; no conflicts were found.

• RRSI is attributed to Google Research in both the Hugging Face report (2026-09-20T20:00:00.000Z) and the arXiv report (2026-09-21T17:54:49.000Z).
• It regularizes recursive self-improvement of agent-harness components: prompts, control flow, tools, memory, and context.
• A temporally annealed edit budget and history-based exploration favoring unexplored evolution paths limit proposed edits.
• A critic rejects benchmark-specific proposals; a pruner drops tiny, costly, or obsolete changes.
• Evaluation spans eight coding, workspace, and engineering-design benchmarks.
• Gains reach up to 14.1 points in-distribution and 4.7 points on five out-of-distribution test sets.
• The resulting harness uses 30% fewer policy tokens than unregularized evolution.
• Code is released under Google Research, per the Hugging Face report.