WIDE: Wildcard Inference for Cross-Modal Generative Retrieval
September 4, 2026
WIDE addresses forced hallucinations in cross-modal generative retrieval caused by information asymmetry between text queries and visual candidates. It employs Adaptive Entropy Thresholding to calibrate uncertainty boundaries, preventing irrelevant candidates from hijacking top rankings during decoding.
HOW THIS AFFECTS YOU
●
builderYou can improve the accuracy of multimodal search engines by reducing decoder hallucinations.
●
researcherYou can leverage this method to handle structural mismatches in sequence-to-sequence cross-modal tasks.