PEAR: Adaptive Advantage Reshaping for Perception and Exploration in Large Vision-Language Models
arXiv:2610.32216v1 Announce Type: new Abstract: Effective multimodal reasoning requires a model to reliably utilize visual information while sufficiently exploring diverse reasoning paths. To achieve this goal, prior work combines visual...
Read more: https://arxiv.org/abs/2610.32216
Scan with your phone
Scan with your phone