Reasoning Models Know What's Important, and Encode It in Their Activations
For researchers studying how language models process reasoning, this work shows that analyzing activations reveals internal representations of step importance that surface-level token analysis misses.
The paper investigates whether model activations or tokens better identify important reasoning steps in language models, finding that activations encode more information and generalize across models, while tokens do not capture step importance as effectively.
Language models often solve complex tasks by generating long reasoning chains, consisting of many steps with varying importance. While some steps are crucial for generating the final answer, others are removable. Determining which steps matter most, and why, remains an open question central to understanding how models process reasoning. We investigate if this question is best approached through model internals or through tokens of the reasoning chain itself. We find that model activations contain more information than tokens for identifying important reasoning steps. Crucially, by training probes on model activations to predict importance, we show that models encode an internal representation of step importance, even prior to the generation of subsequent steps. This internal representation of importance generalizes across models, is distributed across layers, and does not correlate with surface-level features, such as a step's relative position or its length. Our findings suggest that analyzing activations can reveal aspects of reasoning that surface-level approaches fundamentally miss, indicating that reasoning analyses should look into model internals.