Reasoning Models Reason Inefficiently

Dipika Khullar, Ashwinee Panda

Keywords: large language models, reasoning traces, backtracking, efficiency, interpretability, internal activations, steering vectors

TL;DR: We make models more token efficient by steering away the backtracking direction.

Abstract: Large language models (LLMs) produce long, structured reasoning traces that can inflate latency and cost. Our results suggest that while backtracking can help models arrive to the correct answer, they are not a faithful picture of the minimal computation required to solve a task—they can be compressed or restructured. In this paper, we show how to build more efficient and interpretable reasoning processes by identifying and targeting internal directions associated with inefficiency.

Submission Number: 100