State of VLA Research at ICLR 2026 (mbreuss.github.io)

At ICLR 2026, the state of Vision-Language-Action (VLA) models was analyzed, showcasing the rapid growth and evolving definitions within this field. There remains a contentious debate regarding what constitutes a VLA, with varying interpretations regarding the significance of internet-scale pretraining. The author emphasizes that a proper VLA model should be pretrained on vision-language data to enhance its instruction-following capabilities and task generalization. Current challenges, such as zero-shot generalization and performance on complex tasks, suggest that many VLAs still function more as sophisticated multimodal policies than as truly capable robotic agents. The conference submissions highlighted key trends, including discrete diffusion models, which promise faster sequence generation for action commands, and approaches to embody reasoning through embodied chain-of-thought methodologies. The role of new tokenizers for converting continuous actions into discrete, manageable tokens for VLA training is also gaining traction, addressing issues of performance and efficiency. Additionally, the potential incorporation of video prediction models indicates a growing interest in leveraging motion dynamics for improved robotic control. Overall, as the interest in VLAs surges, researchers face exciting challenges in enhancing model robustness and generalization across diverse tasks, setting the stage for further innovation in this transformative area of AI.