Researchers have proposed a new approach to vision-language-action models called LoopVLA. This architecture learns to refine multimodal representations and estimate when they are sufficient for action prediction. Unlike existing methods that rely on predefined layers or heuristic rules, LoopVLA iteratively refines the representation and produces both a candidate action and a sufficiency score. The model is trained using a self-supervised objective that links sufficiency learning to policy optimization signals. Experiments show that LoopVLA can reduce parameters by 45% and improve inference throughput while matching or outperforming strong baselines in task success. --- Why it matters: This matters because it addresses the issue of excessive abstraction in vision-language-action models, which can waste computation and weaken low-level cues essential for precise control. By learning to estimate when a representation is sufficient, LoopVLA can improve the efficiency-performance frontier of VLA policies. Source: https://arxiv.org/abs/2605.09948