Base Models Know How to Reason, Thinking Models Learn When

What do thinking language models learn during training that their base models lack? We first present an unsupervised method that discovers a model's reasoning behaviors by training small Sparse Autoencoders on sentence-level activations of reasoning traces, yielding interpretable reasoning taxonomies. Building on this, we introduce constructive model diffing, which aims to reconstruct the base-to-fine-tuned difference from interpretable components: reasoning mechanisms (category vectors that can induce a reasoning behavior in the base model) and reasoning heuristics (a classifier determining when a mechanism should fire). Across nine base/thinking pairs (four RL-trained, four SFT-distilled, one mixed), two independent findings agree: category vectors in the base model converge to far lower loss for taxonomies derived from purely RL-trained models, and hybrid models recover roughly 76% of the RL base-to-thinking gap but only 11% of the SFT gap. This indicates RL primarily teaches heuristics for orchestrating pre-existing base mechanisms, whereas SFT-distillation installs new ones, offering a new lens on what training paradigms teach, with implications for efficient reasoning-model development.

AI models have recently become better at hard problems like math by thinking out loud, working through a question step by step before answering. These thinking models are built by giving an ordinary model extra training. But what does that training actually teach: brand new reasoning skills, or just a better sense of when to use skills it already had? We built a way to split the gap between an ordinary model and its thinking version into two parts: the reasoning skills, and knowing when to use each one. We then rebuilt a thinking model by adding only these parts back to the ordinary model, and measured how much performance we recovered. We compared two training methods across nine AI models. The answer depends on the method. When models were trained by trial and error, we recovered most of their ability just by teaching the ordinary model when to use skills it already had. When models were instead trained to imitate a larger expert model, this failed, which means that training had taught them genuinely new skills. This clarifies what different training methods teach, and points to cheaper ways to draw out reasoning that base models may already have.