作者位次 #2/6ACL 2026Findings· Findings of ACL
Look and Think: Efficient Multimodal Reasoning via Modality-Decoupled Compression
Xidi Cai · Junhao Zheng · Jingye Li · Boyuan Li · Shaowei Zhang · Qianli Ma
摘要
Multimodal large language models (MLLMs) have achieved strong performance on challenging visual question answering benchmarks, yet their inference efficiency is severely constrained by the rapidly growing context. This growth stems from two primary sources: the large number of visual tokens requi...