Donghyeon Cho, Joowon Kim, Seungho Shin et al. · 27 Oct 2026 · single gpu · code

A training-free decoding rule for diffusion VLMs runs a cheap text-only pass at each step and processes the image only when that pass disagrees with the last image-conditioned prediction, keeping 99.6% of the score at about a third of the FLOPs.

Reproducing it means writing a custom denoising loop for LLaDA-V, an 8B model that fits on a single 24GB GPU in bf16. The loop compares masked-token predictions with and without the image, reuses the image-conditioned predictions that the text-only pass agrees with, and lets the count of agreeing tokens set how many get unmasked. There are clear numbers to check: the 22.8% of steps where removing the image changes a token being unmasked, and the score-versus-FLOPs tradeoff on a few benchmarks. Along the way you learn how masked diffusion decoding works step by step, and how agreement can stand in for a threshold in adaptive compute, an idea that carries over to caching and early exiting elsewhere.

discuss → · arXiv · pdf · hf · alphaXiv · code