LOWN cuts LLaDA-V decoding to 35.3% of FLOPs at 99.6% score

A new method called LOWN routes the image through a diffusion vision-language model only when the model's text prediction actually depends on it, retaining 99.6% of LLaDA-V's score on ten benchmarks while using 35.3% of the decoding FLOPs on average. The paper, posted to arXiv by Donghyeon Cho, Joowon Kim, Seungho Shin and Eunho Yang, requires no training and no tuned threshold.

The authors' starting observation is blunt: visual tokens make up 89-97% of a diffusion VLM's input on the benchmarks they study, and "most of the decoding computation is spent repeatedly processing the same image." When they ran LLaDA-V with and without the image at every decoding step, they found that "removing the image changes a token being unmasked at only 22.8% of steps on average and at as few as 7.9%."

LOWN exploits that gap. At each step it predicts the masked tokens without the image, then processes the image only when that prediction disagrees with the last image-conditioned prediction. "Agreement certifies the earlier prediction for reuse," the paper says, and the number of certified tokens sets how many are unmasked. The authors note that "the need for the image differs across benchmarks," so a fixed schedule would leave accuracy on the table.

LOWN generalizes across LLaDA-based DVLMs, composes orthogonally with visual token pruning, and runs under key-value caching for further gains. The paper reports no wall-clock latency numbers, no per-benchmark accuracy breakdown, and no results on diffusion VLMs outside the LLaDA family.