When Reasoning Helps Action: Monitoring and Steering Chain-of-Thought in Vision-Language-Action Policies
The paper introduces a method to monitor and steer reasoning traces (Chain-of-Thought) in Vision-Language-Action (VLA) policies for runtime safety. They developed TRUST, an offline value model that predicts reasoning correctness. This allows the system to detect unreliable reasoning during generation and use gated sampling to correct it, aiming to improve safety by steering the policy toward better actions.