Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning

The paper addresses a train-inference mismatch in continuous latent reasoning for multimodal large language models. A training-time posterior conditioned on the ground-truth answer can exploit shortcuts unavailable to the inference-time prior, causing answer leakage and prior contamination. The proposed Asymmetric Mutual Variational Learning (AMVL) uses forward KL to align an answer-agnostic prior with the posterior, while a novel reverse KL regularizes the posterior against inference-incompatible regions. In a latent-integrated MLLM, AMVL reportedly improves the complex BLINK benchmark average by 10.83 points, with gains up to 32.00 points on individual tasks.