Peking Team's Harness-Zero Lifts Base LLM From 23.3% to 44.3%

Distilling agent-harness behaviors back into a student model's weights lifts a base LLM's macro-average task success from 23.3% to 44.3% across three benchmarks, and the distilled model beats the 41.7% the same base scores with the specialized harness still bolted on at deployment. That is the headline claim of Harness-Zero, a paper posted to arXiv on September 21 by a team led by Haoran Ye at Peking University.

The method they call agent-as-harness replaces the usual code scaffolding with a reviewer model. A harnessing agent (GPT-5.6 Sol in the paper) inspects the Qwen3.5-9B student's responses inside the target harness's action space, corrects them where needed, and the resulting trajectories become supervised fine-tuning data. "Harness-Zero recovers harness-induced behaviors absent from the base model, with 82.3% average recovery across 28 patterns in the three domains," the abstract reports.

On six settings across SpreadsheetBench Verified, AppWorld and USPTO Retrosynthesis, the agent-as-harness pipeline averages 81.1% for frontier LLMs versus 78.1% for the code-as-harness baseline on the same evolved harness. "For frontier LLMs using the same evolved harness, agent-as-harness outperforms code-as-harness," the authors write.

The gains are not uniform. On USPTO Retrosynthesis, the distilled model lands at 30.0% while the meta-harness at deployment reaches 38.0%, and the authors flag that "procedural harness behavior is easier to internalize than deep domain knowledge." Trajectory collection also runs 2.4x slower than a vanilla mini-SWE-agent trial, though that cost is paid at training time rather than at inference. The preprint lands the same day as a related result on regularizing harness self-improvement, one of several harness-flavored papers moving through the analyst feeds this month.