DynaSteer Framework Guides LLMs Towards Truth in Reasoning

Current methods to enhance Large Language Model (LLM) reasoning often focus on encouraging more thought, but struggle to consistently guide models towards factual accuracy. This paper addresses this gap by exploring how to intrinsically control LLM reasoning trajectories. The researchers uncover key insights: truth is encoded at the sentence level and intertwined with latent reasoning patterns. They also identify an "Uncertainty Principle" and "Decay Effect," suggesting that effective interventions must be localized to early, high-entropy decision points in the reasoning chain to avoid unintended consequences. Based on these findings, the paper proposes DynaSteer, a dynamic representation editing framework. DynaSteer uses pattern clustering to separate reasoning manifolds and Fisher-LDA to project purified truth. By monitoring lookahead entropy, it selectively steers and rolls back trajectories only when necessary, demonstrating effectiveness on MATH benchmarks and generalization to coding tasks.