Fantastic Reasoning Behaviors and Where to Find Them: Unsupervised Discovery of the Reasoning Process

TL;DR

This paper introduces RISE, an unsupervised auto-encoder framework that uncovers and controls diverse reasoning behaviors in large language models by analyzing activation space, enabling interpretability and manipulation of reasoning processes.

Contribution

The work presents a novel unsupervised method using sparse auto-encoders to discover and control reasoning behaviors in LLMs without human-defined concepts.

Findings

Disentangles reasoning behaviors like reflection and backtracking in activation space

Enables controllable amplification or suppression of specific reasoning behaviors

Discovers novel reasoning behaviors beyond human supervision

Abstract

Despite the growing reasoning capabilities of recent large language models (LLMs), their internal mechanisms during the reasoning process remain underexplored. Prior approaches often rely on human-defined concepts (e.g., overthinking, reflection) at the word level to analyze reasoning in a supervised manner. However, such methods are limited, as it is infeasible to capture the full spectrum of potential reasoning behaviors, many of which are difficult to define in token space. In this work, we propose an unsupervised framework (namely, RISE: Reasoning behavior Interpretability via Sparse auto-Encoder) for discovering reasoning vectors, which we define as directions in the activation space that encode distinct reasoning behaviors. By segmenting chain-of-thought traces into sentence-level 'steps' and training sparse auto-encoders (SAEs) on step-level activations, we uncover disentangled features corresponding to interpretable behaviors such as reflection and backtracking. Visualization and clustering analyses show that these behaviors occupy separable regions in the decoder column space. Moreover, targeted interventions on SAE-derived vectors can controllably amplify or suppress specific reasoning behaviors, altering inference trajectories without retraining. Beyond behavior-specific disentanglement, SAEs capture structural properties such as response length, revealing clusters of long versus short reasoning traces. More interestingly, SAEs enable the discovery of novel behaviors beyond human supervision. We demonstrate the ability to control response confidence by identifying confidence-related vectors in the SAE decoder space. These findings underscore the potential of unsupervised latent discovery for both interpreting and controllably steering reasoning in LLMs.