New diagnostic reveals multimodal AI ignores reliability scores

A new paper from Moon, Pillai, and Campbell, accepted to INTERSPEECH 2026, tackles a subtle but critical question in multimodal AI: when a model fuses inputs from different sources (e.g., video and audio) and assigns reliability scores to each modality, does it actually use those scores to inform its final decision? To find out, the researchers designed a simple diagnostic: after training, they permute the reliability scores across test examples. If the model truly depends on them, performance should degrade. Surprisingly, on two standard benchmarks—StressID for stress recognition and CMU-MOSEI for sentiment analysis—permuting the scores left performance virtually unchanged, despite the fact that selecting the best modality per example could yield large gains.

The paper then runs positive controls where the reliability signals are engineered to actually indicate which modality is correct. In those cases, the same frozen fusion rules produce significant performance improvements. This shows that current multimodal models are not inherently ignoring reliability scores—they only leverage them when the scores reliably predict unimodal correctness. The diagnostic is "leakage-safe" because it avoids data contamination in the testing procedure. For practitioners, this work serves as a warning: your multimodal model might be fusing information but not truly weighting it by confidence, leaving potential robustness gains on the table. The paper's method offers a straightforward way to audit any multimodal system for this behavior.