Lichao Sun

We introduce an unsupervised method for post-training multi-modal large language models using implicit reward signals from majority voting based on GRPO.

This paper introduces CorrectBench, the first comprehensive benchmark for systematically evaluating self-correction mechanisms in LLMs.

We conducted the first study on backdoor attacks against vision-Language-Action models, introducing an effective backdoor attack algorithm that achieved a 100% success rate, providing valuable insights for VLA research.