A new interpretability paper (arXiv:2510.07364) argues that the gap between a base language model and a "thinking" model is mostly about when to reason, not whether the model already knows how.
The researchers trained small Sparse Autoencoders on sentence-level activations drawn from model reasoning traces, generating what they call reasoning taxonomies. They then applied a technique called "constructive model diffing" to measure what training changed between a base model and its fine-tuned thinking counterpart. Across nine base/thinking model pairs (four trained with reinforcement learning, four with supervised fine-tuning distillation, and one mixed), two consistent findings emerged: RL-trained model taxonomies converged to far lower loss using base-model category vectors, and hybrid models recovered about 76% of the RL-to-thinking gap but only 11% of the SFT gap.
The split matters because it reframes what "thinking" models actually do differently. If RL training is largely teaching scheduling heuristics rather than new reasoning skills, then benchmarks that pit base models against thinking models may be measuring when the model chooses to reason, not how capable its reasoning is. The SFT-distillation finding cuts the other way: distilling a thinking model apparently installs genuinely new reasoning mechanisms, not just new heuristics for deploying old ones, which makes the SFT gap harder to close than the RL gap.
The paper is already on version four, which suggests it has faced meaningful scrutiny; treat the sparse autoencoder methodology as an active proposal in the interpretability community, not a settled technique.