arxivcs.CV2026-07-01
Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning
Shijie Li, Yilin Gao, Siyuan Yang, Tieyuan Chen, Chaofan Gan, Zhihao He, et al.
Multimodal Large Language Models (MLLMs) are often constrained by a language-space bottleneck, forcing complex visual reasoning into discrete tokens which can lose perceptual nuance. A promising alternative is continuous latent reasoning, where the goal is to discover implicit re…