Self-Harness: Harnesses That Improve Themselves

Overview

Self-Harness keeps the model weights and evaluator fixed while improving the surrounding harness. Each round evaluates the current harness on tasks, mines failure patterns from execution traces, asks the same model to propose bounded harness edits, and promotes a candidate only when held-in and held-out regression checks support the change.

Results

Terminal-Bench-2.0 pass rates (%). More results are coming soon.

Model	Initial	Self-Harness
MiniMax M2.5	42.2	53.9
Qwen3.5-35B-A3B	18.0	36.7
GLM-5	46.1	57.0

Citation

@misc{zhang2026selfharnessharnessesimprove,
      title={Self-Harness: Harnesses That Improve Themselves},
      author={Hangfan Zhang and Shao Zhang and Kangcong Li and Chen Zhang and Yang Chen and Yiqun Zhang and Lei Bai and Shuyue Hu},
      year={2026},
      eprint={2606.09498},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.09498},
}