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

Short summary

DynaSteer, a dynamic representation editing framework, improves LLM reasoning by identifying and steering critical decision points in reasoning chains. Using pattern clustering and Fisher-LDA, it isolates truth-encoding neurons and intervenes at high-uncertainty forks while avoiding collateral damage. Comprehensive validation on MATH and coding benchmarks with open-source code available.

• •Dynamic representation editing framework for steering LLM reasoning trajectories
• •Identifies truth patterns and high-entropy decision points for selective intervention
• •Validated on MATH benchmarks and out-of-domain coding tasks with public code release

Generated with AI, which can make mistakes.