Search for Truth from Reasoning: A Dynamic Representation Editing Framework for Steering LLM Trajectories

2026-06-26

AI Summary

This work addresses the challenge that current large language models lack precise mechanisms to dynamically intervene in reasoning trajectories, often failing to steer inference effectively toward correct answers. To overcome this limitation, the authors propose DynaSteer, a framework that analyzes the geometric structure of "truthful" reasoning paths and applies selective intervention with trajectory rollback at high-entropy early branching points. DynaSteer integrates dynamic representation editing (RepE), lookahead entropy monitoring, pattern clustering, and Fisher-LDA–based purification of truthful vectors to preserve valid reasoning pathways while correcting deviations. Experimental results demonstrate that DynaSteer significantly improves accuracy on multiple MATH benchmarks and exhibits strong out-of-domain generalization on programming tasks.

Abstract

Current approaches to enhance Large Language Model (LLM) reasoning, such as Chain-of-Thought and "Wait" prompts, primarily encourage models to think more, yet often fail to guide them toward Truth. While Representation Editing (RepE) offers a intrinsic control, its application to dynamic reasoning trajectories remains underexplored. In this work, we bridge this gap by investigating the geometry of truth within unfolding reasoning chains. We uncover three critical insights: (1) Truth is encoded at the sentence level and is entangled with latent reasoning patterns; (2) Effective intervention follows an Uncertainty Principle and a Decay Effect, requiring localization to early, high-entropy forks; (3) Naive steering vectors suffer from noise, risking collateral damage to correct trajectories. Based on these findings, we propose DynaSteer, a dynamic RepE framework. DynaSteer employs pattern clustering to disentangle reasoning manifolds and utilizes Fisher-LDA to project purified truth. By dynamically monitoring lookahead entropy, it selectively steers and rolls back trajectories only when necessary. Comprehensive experimental results on several MATH benchmark verify the effectiveness of DynaSteer, and experiments on out-of-domain coding tasks further confirm its generalization ability. Our code is publicly available at https://github.com/tianlwang/DynaSteer.

Problem

• Large Language Models
• Reasoning
• Truth
• Representation Editing
• Dynamic Steering

Innovation

• Dynamic Representation Editing
• Reasoning Trajectory Steering
• Truth Geometry
• Fisher-LDA
• Lookahead Entropy