Embodied agents act better when they decide before explaining
Where Do Embodied Decisions Come From? Rethinking Latent and Explicit Reasoning
Computational Engineering, Finance, and Science
Summary
Sometimes robots or AI models that see and act make worse decisions when they try to explain their reasoning step by step before acting. The authors show that if these agents choose their action first and then explain why, they perform better across tasks like driving or manipulating objects. This new way, called decide-then-explain, relies more on directly using what the agents perceive instead of heavy explicit reasoning before acting. The authors suggest using explanations mainly during training to improve decisions rather than at the moment of action.
What this means in practice
- •For autonomous vehicle teams: Improve autonomous driving by adopting a decision-before-explanation approach to increase navigation accuracy and robustness in real-time.
- •For robotic system developers: Build robotic manipulators that prioritize action prediction before reasoning outputs to enhance manipulation performance across various tasks.
Authors
Yuan Lin, Ziyue Zhou, JinLong Zhao, Pei Liu, Haipeng Liu, Pan Zhou, Kun Zhan
Abstract
Chain-of-Thought (CoT) reasoning is increasingly incorporated into Vision-Language-Action (VLA) models, yet it can degrade the performance of stronger embodied agents. We investigate this capability-dependent effect by distinguishing explicit reasoning from latent decision computation, i.e., perception-grounded computation that directly supports action prediction. Under the standard \textit{think-then-act} (TTA) paradigm, intervening on the generated CoT while fixing the visual input and model parameters causes a substantial performance collapse, with Navigation F1 dropping from 72.14% to 11.84%, demonstrating the strong influence of explicit reasoning on action generation. We then propose \textit{decide-then-explain} (DTE), which predicts actions before generating explanations, and introduce Visual Conditional Contribution (VCC) and Reasoning Conditional Contribution (RCC) to characterize the resulting decision process. Across autonomous driving and robotic manipulation, DTE consistently outperforms TTA and conventional \textit{no-CoT} baselines, while exhibiting greater reliance on perception-grounded computation. Further TTA-trained, DTE-inference experiments show that this benefit is not solely attributable to retraining under the new factorization. Our results suggest that for capable embodied agents, explicit CoT may be better used to shape decision computation during training rather than mediate action generation at inference time. Code: https://github.com/ocean-luna/openvla-decide-then-explain.