Perception or Reasoning? Diagnosing and Tackling Reinforcement Learning Bottlenecks in VLMs
GRAD, a group-relative attention distillation objective that uses the best rollout in each group to directly guide the visual attention of weaker rollouts is proposed, which consistently improves performance across two model scales and seven benchmarks, and further complements existing RL methods.
Ze-Yuan Yang, Zhuo-Hao Chen, Jia-Ben Chen et al.
· 0 citations