Self-HPO alternates harness and policy updates for coding agents
Original title: Self-Evolving Coding Agents via Harness-Policy Optimization
On Qwen3.5-35B-A3B, average resolve rate rose from 57.8% to 64.8% across three coding benchmarks.
Self-HPO is a recursive framework that alternates harness proposals from the current policy with reinforcement learning under the selected harness. Hard tasks found in training become the next harness search set so updates track the policy’s failure modes. Across three coding benchmarks it beat harness-only and policy-only baselines, and joint gains exceeded the sum of each component’s gains. A stronger external harness proposer such as Opus 4.8 gave a further boost.
Unconfirmed: Benchmark names are not given, so the 64.8% figure cannot be checked against a named suite.
- Qwen3.5-35B-A3B average resolve rate moved from 57.8% to 64.8%.
- Each round has an H-step for harness candidates and a P-step of RL.
- Both optimized harness and policy kept gains when paired with base counterparts.
- Pass@1 reliability and Pass@k capability frontier both improved.
- Using Opus 4.8 as harness proposer added a further lift.
From the original article
1 min read at huggingface.co
Large language model (LLM) agents have shown promising potential in coding tasks, and their performance is jointly determined by the policy model and the harness that mediates their interaction with the environment.
By Zhepei Wei, Wenlin Yao, Baolin Peng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.