Awesome Zipeng Fu!! 👏 Love the demos, and great to see that everything is open-source!

Can we train VLAs to think about what to do next—visually—before executing tasks? In this work led by Qingqing Zhao, we found that *visual* chain-of-thought reasoning enhances policy success rates + enables VLAs to leverage unlabeled video data during pretraining! #CVPR2025

RL with verifiable reward has shown impressive results in improving LLM reasoning, but what can we do when we do not have ground truth answers? Introducing Self-Rewarding Training (SRT): where language models provide their own reward for RL training! 🧵 1/n