In self-evolving frameworks, task solutions are typically optimized while treating the surrounding harness as fixed. We introduce Harness-Aware Self-Evolving (HASE), an agentic reinforcement-learning framework that allows a single model to 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. Additionally, HASE repairs imperfect evaluation components and converges to state-of-the-art performance in circle-packing algorithm discovery. These results show that HASE improves both the harness and the solution through one unified agentic process.
Blogger's Review: The innovation of the HASE framework lies in its integration of harness and task solution optimization, showcasing a new approach in the field of reinforcement learning. Its performance surpassing traditional large models marks a significant advancement in self-evolving technology, warranting further exploration and application.