TAD: Token-Adaptive Contrastive Decoding with Confidence-Guided Gating for Hallucination Mitigation in Large Audio-Language Models
Token-Adaptive Decoding is proposed, a training-free strategy for hallucination mitigation that grounds the initial yes/no decision by contrasting logits under real audio with a matched silent reference, and introduces a token-adaptive, confidence-guided gate that is decision-critical at the first decoding step.
He-Yu Chang, Nianwen Si, Hao Zhang et al.
· 1 citation
· ⚡1