Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs

A research paper proposes certificate-guided visual routing for diffusion vision-language models (DVLMs), which process their full input at every denoising step even though visual tokens account for 89-97% of that input on the benchmarks the authors studied. The method targets the repeated re-processing of the same image during decoding, an inefficiency that existing visual token pruning methods only partially relieve.

Diffusion vision-language models DVLMs process their entire input at every denoising step. Since visual tokens make up 89-97% of this input on the benchmarks we study, most of the decoding computation is spent repeatedly processing the same image. Visual token pruning methods partially relieve thi