ASCD: Attention-Steerable Contrastive Decoding for Reducing Hallucination in MLLM
ASCD:用于减少多模态大语言模型(MLLM)幻觉的注意力可导向对比解码
专题命中 跨模态检索 :MLLM(title,title_cn);multimodal(abstract);cross-modal(abstract);分类 cs.CV、cs.CL
AI总结 该研究针对MLLM的幻觉问题,提出ASCD方法,通过正负引导调整解码时的注意力分数,在多基准上显著减少幻觉并提升VQA准确率,且无需额外训练。
Comments Accepted at AAAI 2026
Journal ref Proceedings of the AAAI Conference on Artificial Intelligence, 40(12): 10306-10314, 2026