SynIB: Informational Bottleneck for Maximizing Synergy in Multimodal Learning

In three linesSynIB, a new information-theoretic training objective, maximizes multimodal synergy by penalizing model confidence when any modality is masked. Validated on synthetic XOR and five real benchmarks (MultiBench, Hateful Memes, CREMA-D), SynIB improves accuracy on synergy-dependent examples by up to 7.8% and overall accuracy by 3.8%.

Summary generated by Claude — human-verified