NoThink Post-Training Gains Are 42-79% 'Thinking Leakage' From the Base Model's Think Mode

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.28682 audits hybrid reasoning models post-trained in NoThink mode with a causal mediation framework. Steering the base model along one activation direction reproduces most of the post-training gain on competition math. Counter-steering a checkpoint removes much of what it gained. Across nine checkpoints the leakage ratio was 42-79%, so a NoThink method that looks better may simply trigger more hidden Think behavior, with the latency cost that implies.