ClawLoop: feedback closes the training cycle

TL;DR

ClawLoop alternates task synthesis, training, and diagnosis to improve a harness-based agent without an external teacher model. The abstract reports benchmark gains but gives no numerical effect sizes. (source)

Authors & Org

Hansi Yang, Linji Hao, Hongyan Hao, An Wang, Han Wu, Qi Gu and Xunliang Cai. Submitted 2026-10-09; affiliations are not established by the captured abstract page. (source)

Method

Task synthesis combines two construction routines, diagnosis-conditioned difficulty growth, and validation both with and without an agent. SFT and/or RL then update the agent. A report analyzing rollout behavior guides the next task-synthesis iteration. (source)

Results

The authors report consistent improvements for Qwen3.6-27B on SkillsBench, Claw-Eval-Live and ClawGym-Bench. Refined tasks also improve Qwen3.5-4B and 9B across the benchmarks. The abstract supplies neither scores nor training-cost comparisons. (source)

Significance

For Agentic Reinforcement Learning, evaluation feedback becomes an input to the next training cycle rather than a report consumed only by researchers. The claim concerns a specific iterative pipeline, not unrestricted recursive self-improvement. (source)

Open Questions

How much do task validation and diagnosis contribute separately, and do gains survive matched total compute? The abstract does not answer these questions; the full paper and code have not been reviewed here. (source)

Cite

• ClawLoop: Self-Improving Agents via Harness-Native Training Loop, arXiv:2610.14118.
• Captured abstract page.