Reflective Recovery Enables Self-Supervised Learning from Reasoning Mistakes

Reflective Recovery addresses scaling collapse in reasoning tasks by using failed trajectories as training data. The method extracts initial segments of failed attempts and uses them as prompts to guide the model toward valid solutions, effectively teaching the model to recover from errors.