Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs
LOWN retains 99.6% of LLaDA-V’s benchmark score while using 35.3% of its decoding FLOPs, the authors report.
Across ten benchmarks, visual tokens account for 89–97% of the input that diffusion vision-language models repeatedly process. LOWN first predicts without the image, processing it only when that prediction disagrees with the last image-conditioned prediction. Agreement allows earlier predictions to be reused and determines how many tokens are unmasked at that step. The method requires no training or tuned threshold and can be combined with visual token pruning and key-value caching.
HF Daily Papers' note
Across ten benchmarks, visual tokens account for 89–97% of the input that diffusion vision-language models repeatedly process. LOWN first predicts without the image, processing it only when that prediction disagrees with the last image-conditioned prediction. Agreement allows earlier predictions to be reused and determines how many tokens are unmasked at that step. The method requires no training or tuned threshold and can be combined with visual token pruning and key-value caching.
HF Daily Papers' note
score 5