Overview
Field: NLP Authors: Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen, Yufan Zhang, Yoonho Lee, Chengsong Huang, Han Yu, Zhongying CuiZhu, Yifei Ming, Huaxiu Yao, Burak Gokturk, Tomas Pfister, Chen-Yu Lee Published: 2026-09-21 arXiv: 2609.24972
Abstract
An LLM agent's capability is largely magnified by its harness, namely the prompts, control flow, tooling, memory, and context management surrounding the frozen backbone model. Recent methods increasingly automate this process by iteratively proposing and selecting component-wise edits of an agent harness, practically establishing a form of recursive self-improvement (RSI) at the agent-system level. However, such recursive evolution may overfit by memorizing the training tasks, showing large in-distribution gains that shrink or even vanish on out-of-distribution benchmarks.
The authors introduce RRSI (Regularized Recursive Self-Improvement of Agent Harnesses), which incorporates regularization principles into harness self-improvement by constraining both candidate proposal and selection:
- Proposer: applies a temporally annealing budget that limits how many edits candidates can bundle, and encourages exploration of unexplored paths based on evolution history.
- Selector: equipped with a critic that filters benchmark-specific proposals and a pruner that removes too-small, too-expensive, or no-longer-useful changes.
- Up to 14.1 points improvement on evolution-targeted (in-distribution) splits
- Up to 4.7 points improvement on five out-of-distribution benchmarks
- 30% reduction in policy token consumption compared to unregularized evolution
Together, these constraints favor reusable agent mechanisms over benchmark-specific ones or noise.
Results
Across eight benchmarks spanning coding, agentic workspaces, and engineering design tasks:
*Auto-collected on 2026-09-23*