Tsinghua, Zhipu and partners introduced HarnessPro, which lets agents study their own runs and rewrite their harness without reference answers.

The team reports a 2-point pass-rate gain and a 26.1% cut in cost per task, part of a broader Chinese push into recursive self-improvement around model scaffolding rather than weights.

AI-generated: summaries written by AI from the linked sources. How we use AI