AMVL Framework for Continuous Multimodal Latent Reasoning

Asymmetric Mutual Variational Learning (AMVL) addresses the train-inference mismatch in multimodal continuous reasoning. It prevents models from exploiting answer-dependent shortcuts during training, allowing for more robust latent reasoning pathways that do not rely on ground-truth leakage.