Paper summary
Paper summary: Reflect, Retry, Reward: Self-Improving LLMs via Reinforcement Learning
When a model fails a task, it writes a self-reflection and tries again; if the retry succeeds, the reflection tokens are rewarded, which trains better self-reflection with only binary feedback.
- Only the reflection is rewarded, not the answer, so the model learns to write reflections that actually help.
- The paper reports fine-tuned 1.5B–7B models outperforming same-family models ten times larger.