Negative Self-Distillation: Learning to Reason by Avoiding Flaws

Standard on-policy self-distillation often degrades complex reasoning by forcing models to mimic artificially confident traces, suppressing the uncertainty and exploratory backtracking essential for hard problems. Negative Self-Distillation (NSD) inverts this paradigm by optimizing models to steer clear of their own flawed reasoning rather than imitating ground-truth solutions. To avoid degrading general language skills during unlearning, NSD uses a dynamic gating mechanism that isolates reasoning-critical tokens from routine linguistic priors. The framework consistently outpaces conventional self-distillation and label-free reinforcement learning baselines.