Research on Self-Repair Retry Mechanisms for Code-as-Policy Agents
What happened
Siddharth Aphale and Ayushman Singh submitted a paper studying retry mechanisms for self-repair in code-as-policy agents on simulated robotic manipulation tasks. On the same 252 initial failure cases, Qwen2.5-Coder-7B, using feedback templates selected by failure label, recovered 85.3% within five retries; under the same retry limit, resampling recovered 68.3%. Under a fixed total token budget, this advantage persists. The study also found that template feedback is not always effective: on a nut-insertion task requiring edge-grasping, template feedback instead reduced the recovery rate, which the authors attribute to the template giving incompatible center-grasp instructions.
AI 根据报道生成 · 6 days ago更新
Reporting timeline
Follow the reporting to explore different perspectives.
- arXiv cs.ROCode as Policy Agents' Self-Repair: What Drives Retry Success
This study compared retry mechanisms of code-as-policy agents on simulated robot manipulation tasks: on the same 252 initial failure cases, Qwen2.5-Coder-7B, using feedback templates selected by failure labels, recovered 85.3% within five retries, versus 68.3% for resampling with the same retry cap, and the advantage persisted under a fixed total token budget.
Activity for this event
- Current
- 3
- Peak
- 10Nov 5 14:00
- Past 24 hours
- -49%
The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.