VAttnPO: Visual Attention-Guided Policy Optimization
This work introduces VAttnPO (Visual Attention-Guided Policy Optimization), which uses answer-conditioned visual attention to refine token-level credit while preserving the verifier-derived trajectory-mean advantage to the GRPO sequence advantage.
Li-Hao Zheng, Zhen-Wei Shao, Zhou Yu et al.
· 0 citations