Negative Self-Distillation trains a model by pushing away from a self-generated 'careless reasoner' instead of imitating solutions

arXiv (2609.11699)·medium signalXBlueskyLinkedInCopy link

The paper starts from the finding that On-Policy Self-Distillation degrades hard-reasoning performance because conditioning on ground-truth solutions produces artificially confident traces, which suppresses hedging and penalizes the exploratory self-correction that hard problems need. NSD inverts the setup: the model generates a question-specific negative condition, such as acting as a careless reasoner, and the student distribution is pushed away from that self-generated negative teacher. No ground-truth answers or external supervision are required, which matters for domains where you have problems but no verified solutions.