DREvo: Distilling Recalibrated Historical Experience for Harness Self-Evolution

2026-07-29

AI Summary

This work addresses the instability and performance fluctuations in existing self-evolution methods for large language model agents, which stem from the lack of dynamic evaluation of historical experience validity and clear guidance for search direction. To overcome these limitations, the paper proposes DREvo, a novel approach that introduces, for the first time, function-level evidence anchoring, state-dependent evidence recalibration, and role-conditioned search intent distillation. These mechanisms enable dynamic filtering of historical experiences and precise generation of evolutionary trajectories. Built upon an iterative optimization framework grounded in historical experience, DREvo achieves state-of-the-art accuracy across five benchmarks, outperforming baseline methods by an average of 16.2% on domain reasoning tasks and 14.2% on agent-based tasks, thereby substantially enhancing both the stability and efficiency of the evolution process.

Abstract

Harness plays a critical role in large language model agent performance, and building a high-performing harness requires substantial expert effort. Therefore, recent research has increasingly explored harness self-evolution, which iteratively proposes, evaluates, and improves harnesses using historical trial experience. However, accumulated historical experience does not always translate into stable search guidance, and performance often fluctuates substantially across evolution iterations, making it difficult to reliably discover high-performing harnesses under a limited evolution budget. We identify two limitations in how existing harness self-evolution methods leverage historical experience: (1) Lack of dynamic reassessment of whether historical experience remains valid for the current harness, and (2) Lack of explicit mechanisms for translating valid historical experience into actionable search directions. To address these limitations, we propose a new harness self-evolution method, named DREvo, which integrates function-level evidence anchoring, state-dependent evidence recalibration, and role-conditioned search intent distillation to determine which historical evidence remains valid and where the harness should evolve next. Under limited evolution budgets, DREvo exhibits smoother evolution trajectories, achieves the highest accuracy on all five benchmarks, and delivers average gains of 16.2% and 14.2% over the evaluated baselines on domain reasoning and agentic tasks, respectively.

Problem

• harness self-evolution
• historical experience
• evolution stability
• search guidance
• limited evolution budget

Innovation

• harness self-evolution
• historical experience distillation
• evidence recalibration
• search intent distillation
• large language model agents