Latent reasoning has advanced multimodal reasoning through a two-stage training paradigm: (1) a helper image is encoded into latent tokens to teach visual ch…
机构:DeepMind
来源:arXiv 2608.19669 | AI4Papers 论文推荐平台