Differentiable Belief-Function Layers Enable Robust Uncertainty Estimation

Replacing Dempster's rule with Yager's rule in deep evidential pipelines prevents gradient divergence when evidence sources disagree. The resulting networks produce an ignorance signal that naturally correlates with human disagreement across visual and linguistic modalities.