Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Zhenyu, Wu, Xiaoxia, Zhou, Zhongzhu, Wu, Qingyang, Zhang, Yineng, Ponnusamy, Pragaash, Subbaraj, Harikaran, Wang, Jue, Song, Shuaiwen Leon, Athiwaratkun, Ben
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917211006828544
author Zhang, Zhenyu
Wu, Xiaoxia
Zhou, Zhongzhu
Wu, Qingyang
Zhang, Yineng
Ponnusamy, Pragaash
Subbaraj, Harikaran
Wang, Jue
Song, Shuaiwen Leon
Athiwaratkun, Ben
author_facet Zhang, Zhenyu
Wu, Xiaoxia
Zhou, Zhongzhu
Wu, Qingyang
Zhang, Yineng
Ponnusamy, Pragaash
Subbaraj, Harikaran
Wang, Jue
Song, Shuaiwen Leon
Athiwaratkun, Ben
contents Large Language Models (LLMs) often rely on long chain-of-thought (CoT) reasoning to solve complex tasks. While effective, these trajectories are frequently inefficient, leading to high latency from excessive token generation, or unstable reasoning that alternates between underthinking (shallow, inconsistent steps) and overthinking (repetitive, verbose reasoning). In this work, we study the structure of reasoning trajectories and uncover specialized attention heads that correlate with distinct cognitive behaviors such as verification and backtracking. By lightly intervening on these heads at inference time, we can steer the model away from inefficient modes. Building on this insight, we propose CREST, a training-free method for Cognitive REasoning Steering at Test-time. CREST has two components: (1) an offline calibration step that identifies cognitive heads and derives head-specific steering vectors, and (2) an inference-time procedure that rotates hidden representations to suppress components along those vectors. CREST adaptively suppresses unproductive reasoning behaviors, yielding both higher accuracy and lower computational cost. Across diverse reasoning benchmarks and models, CREST improves accuracy by up to 17.5% while reducing token usage by 37.6%, offering a simple and effective pathway to faster, more reliable LLM reasoning.
format Preprint
id arxiv_https___arxiv_org_abs_2512_24574
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time
Zhang, Zhenyu
Wu, Xiaoxia
Zhou, Zhongzhu
Wu, Qingyang
Zhang, Yineng
Ponnusamy, Pragaash
Subbaraj, Harikaran
Wang, Jue
Song, Shuaiwen Leon
Athiwaratkun, Ben
Computation and Language
Artificial Intelligence
Machine Learning
Large Language Models (LLMs) often rely on long chain-of-thought (CoT) reasoning to solve complex tasks. While effective, these trajectories are frequently inefficient, leading to high latency from excessive token generation, or unstable reasoning that alternates between underthinking (shallow, inconsistent steps) and overthinking (repetitive, verbose reasoning). In this work, we study the structure of reasoning trajectories and uncover specialized attention heads that correlate with distinct cognitive behaviors such as verification and backtracking. By lightly intervening on these heads at inference time, we can steer the model away from inefficient modes. Building on this insight, we propose CREST, a training-free method for Cognitive REasoning Steering at Test-time. CREST has two components: (1) an offline calibration step that identifies cognitive heads and derives head-specific steering vectors, and (2) an inference-time procedure that rotates hidden representations to suppress components along those vectors. CREST adaptively suppresses unproductive reasoning behaviors, yielding both higher accuracy and lower computational cost. Across diverse reasoning benchmarks and models, CREST improves accuracy by up to 17.5% while reducing token usage by 37.6%, offering a simple and effective pathway to faster, more reliable LLM reasoning.
title Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2512.24574