LOWN released
AIMade by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made
LOWN released — a decoding method that cuts the waste of a diffusion vision-language model reprocessing the whole image at every denoising step: it first tries predicting the masked tokens without the image and processes the image only when that result disagrees with the previous image-conditioned prediction. When the two predictions agree it reuses the earlier prediction, and the number of agreeing tokens sets how many tokens to unmask at that step. arXiv 2610.36671
Why it matters
By the authors' measurement, visual tokens take up 89–97% of the input, yet dropping the image changes which tokens get unmasked in only 22.8% of steps on average, and as few as 7.9%. They reported keeping 99.6% of LLaDA-V's score while using an average of 35.3% of the decoding compute, with neither training nor threshold tuning needed. The claim is that the cost of a diffusion model turns on when it looks at the image rather than on the number of steps, and they wrote that it can be layered on top of visual token pruning or a key-value cache
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2610.36671Added in Week 44, 2026 · See that week
On the map — ㉔ Diffusion Language Models line
- Previous stopResearchers propose register tokens to pass reasoning state in diffusion LMs
- This record · stopLOWN
- Next stopNo later stop yet