Multimodal Continuous Reasoning via Asymmetric Mutual Variational Learning

Researchers introduced Asymmetric Mutual Variational Learning (AMVL) to improve multimodal reasoning by addressing train-inference mismatch. AMVL uses bidirectional calibration to prevent answer leakage and improve latent-space stability, achieving a +10.83 average score improvement on the BLINK benchmark.

1sec.ai