Revisiting the Reasoning Steering Vector: A Truncation Effect

Reasoning-tuned LLMs overthink, and a growing family of efficient-reasoning methods aims to cut the excess. SEAL, a representative approach, steers the residual stream away from reflection and toward execution, reporting higher accuracy at fewer tokens and attributing the gain to modulating reasoning itself. We revisit this reasoning steering vector by decomposing accuracy per problem rather than in aggregate. On MATH-500 and GSM8K, most of the base model's errors are not wrong answers but runaway traces: the model exhausts the token budget without emitting block, so the response is scored wrong even when the correct answer has already appeared. SEAL's improvement is almost entirely the rescue of these truncations rather than the correction of faulty reasoning, and it does not eliminate the failure mode it exploits: it still truncates, and over-steering induces truncations of its own. Accuracy@1 and token count therefore conflate stopping in time with reasoning correctly; efficient-reasoning gains should be measured, and steering methods validated, against that distinction.