Context-Aware Multimodal Claim Verification in Spoken Dialogues

In three linesMAD2 is a benchmark of 1,000 two-speaker audio dialogues (3,368 verifiable claims, ~10h audio) for misinformation detection in conversation. Authors propose calibrated multimodal fusion combining context-aware audio encoder and dialogue-aware text model. Conversational structure improves verification more than misinformation framing.

Summary generated by Claude — human-verified