OpenClaw-RL Improves Self-Distillation Stability Using Text Feedback Overlap

Text feedback is much more informative than outcome in long trajectory tasks! The keys to improve stability and efficiency of self-distillation are found by OpenClaw-RL: 1. Sample Multiple Text Feedback and Select the One that Maximizes Overlap between Teacher and Student! 2. Add Clipping Constant to log p Difference!