Harness-Aware Self-Evolving: Co-Evolving Model Weights, Harness, and Task Solutions

2026-07-04

AI Summary

This work proposes HASE, a novel self-evolution framework that overcomes the limitations of traditional approaches by treating the peripheral harness not as a static component but as an integral part of the co-evolutionary process. For the first time, HASE enables unified co-evolution of model weights, harness components, and task solutions through a Qwen3-8B-based reinforcement learning agent that dynamically edits the harness and jointly optimizes the solution space across multiple rounds of action. By discarding the long-standing assumption of a fixed harness, the method achieves performance on par with the GPT-OSS-120B+Claude Code pipeline on text classification tasks and establishes new state-of-the-art results in both alpha factor mining and circle packing algorithm discovery.

Abstract

Self-evolving frameworks usually optimize task solutions while treating the surrounding harness as fixed. We introduce Harness-Aware Self-Evolving (HASE), an agentic reinforcement-learning framework in which a single model can generate task solutions or edit selected harness components in a multi-turn action space. HASE enables a single Qwen3-8B model to match the text-classification performance of a GPT-OSS-120B model that uses Claude Code as the harness proposer. In alpha factor mining, HASE outperforms the reported GPT-OSS-120B baseline. HASE also repairs imperfect evaluation components and converges to state-of-the-art performance in circle-packing algorithm discovery. These results show that HASE improves the harness and the solution through one unified agentic process.

Problem

• self-evolving
• harness
• task solutions
• co-evolution
• reinforcement learning

Innovation

• Harness-Aware Self-Evolving
• agentic reinforcement learning
• co-evolution
• harness optimization
• Qwen3-8B