Reproduce “Think Less, Act Early: Reinforced Latent Reasoning with Early Exit in Vision-Language-Action Models” on OpenResearch

About this paper

Dianqiao Lei, Lianlei Shan

AVA-VLA gives a vision-language-action robot a learned latent state that updates through task-reward-shaped internal reasoning and stops through a confidence gate before acting. This replaces explicit verbal or visual intermediate plans with adaptive computation. On LIBERO, one policy across four suites averages 98.3% success versus 96.8% for OpenVLA-OFT, while latency gains are measured only in the reported GPU setup.

Key insights

• Implicit latent reasoning, when effectively constrained by task-driven reinforcement learning, can achieve high decision quality and stability without the computational burden of explicit textual reasoning.
• Adaptive computation through an Early-Exit mechanism allows VLA models to dynamically balance reasoning depth with inference efficiency, preventing 'overthinking' for simple tasks and ensuring sufficient deliberation for complex ones.
• Reinforcement learning-optimized latent states can capture abstract, embodiment-agnostic task semantics, leading to enhanced robustness against perturbations and improved zero-shot transfer capabilities across different robotic platforms.

Summary by alphaXiv.