Visual Token Efficiency
Techniques for reducing the number of visual tokens processed by multimodal language models to lower inference cost while maintaining accuracy.
4 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · University of Science and Technology of China, Alibaba Token Hub, Alibaba Group
Adaptive visual tokens cut multimodal inference costs while preserving accuracy
VisionWeave adaptively allocates visual tokens at different spatial granularities, cutting inference costs while preserving accuracy.
Oct 2026 · City University of Hong Kong, Zhejiang University
Adaptive visual memory cuts multimodal reasoning costs without losing accuracy
Reduces multimodal reasoning costs through adaptive visual memory with a low active token budget, improving visual token efficiency.
1 further paper
Oct 2026 · Sun Yat-sen University, Shenzhen Loop Area Institute
Multimodal models learn when to look and how long to reason
this paper proposes dynamic allocation of visual processing by converting intermediate visual tool outputs into tokens, reducing unnecessary visual token usage.
3 of 4 papers shown