Researchers have proposed a new framework called MPCoT for improving the performance of Vision-Language-Action (VLA) policies in long-horizon and high-uncertainty control tasks. The framework uses multi-path latent reasoning to refine multiple hypotheses before action decoding, and is trained using a combination of expert-trajectory consistency, progress scoring, and endpoint-success feedback. MPCoT preserves the original 8-step action interface and generates zero additional tokens during inference, making it more efficient than previous methods. --- Why it matters: This matters because VLA policies are often brittle in long-horizon tasks, limiting their applicability to real-world scenarios. MPCoT's ability to improve performance in these tasks could have significant implications for areas such as robotics and autonomous systems. Source: https://arxiv.org/abs/2606.06245