COOLANT++: concept-bank alignment and confidence-gated reasoning for multimodal fake news detection

Multimodal fake-news detection is complicated by noisy, loosely related, or deliberately mismatched image–text pairs. Existing systems often perform instance-level alignment and evidence aggregation as separate operations, so an alignment score need not directly influence the final modality weights. We present COOLANT++, a framework that represents visual and textual embeddings as distributions over a shared bank of learned prototypes and uses their agreement as an input to a confidence-gated reasoning unit. The model jointly optimizes classification, cross-modal distribution consistency, prototype reconstruction, and gate regularization. On the Twitter and Weibo benchmarks, the submitted experiments report accuracies of 0.900 and 0.923, respectively, with lower performance when either concept alignment or gated reasoning is removed. These results support the proposed coupling on the two evaluated benchmarks; they do not establish that the learned prototypes are human-interpretable concepts or that the model generalizes to unseen events or platforms.