Self-ReSET: RL Framework for Safe Reasoning Recovery

A new reinforcement learning framework called Self-ReSET aims to equip Large Reasoning Models with the ability to self-recover from unsafe reasoning trajectories. Large Reasoning Models excel at self-correction in general domains but often fail to recover from unsafe reasoning paths under adversarial attacks. Existing alignment methods fine-tune models on static expert data with reflection traces or adversarial prefixes, but these approaches struggle because the static training data deviates from the model's dynamic, on-policy reasoning traces. Self-ReSET uses pure reinforcement learning, reusing the model's own safety error trajectories as initial states for training. Experiments across various LRMs and benchmarks demonstrate the framework's effectiveness. The paper is available on arXiv under ID 2605.08936.