Diffusion vision-language models (DVLMs) ship with a cost structure everyone has quietly accepted: every denoising step reprocesses the entire input, and the image dominates it. On October 27, researchers from KAIST and AITRICS posted a paper, Look Only When Needed (arXiv:2610.36671), that first audits this structure: across the ten benchmarks they study, visual tokens make up 89-97% of LLaDA-V's input — yet remove the image entirely, and only 22.8% of decoding steps on average change which token gets unmasked. The low is 7.9% on MMBench; the high, 43.8% on DocVQA. Most denoising steps never needed the image, but the model pays for looking at every single one.

Agreement certifies; disagreement triggers a look

LOWN trains nothing and tunes no threshold. At each denoising step it first runs a cheap TEXT-ONLY forward and compares the result with the last image-conditioned prediction: positions where they agree are "certified" and reused, and the number of certified tokens decides how many get unmasked this step. Only on disagreement does it run a FULL, image-conditioned forward and refresh the anchor. The paper's measurements back the design: FULL and TEXT-ONLY forwards give the same top-1 token at 83.4% of unmasked positions, and 85.0% of final tokens were already the top-1 prediction before being unmasked. Certification has reliability limits — with a one-step-old anchor the disagreement rate is 4.1-11.0%, growing to 25.3% on DocVQA when the anchor is eight steps old — but in practice 92-97% of reused tokens rely on the previous step's anchor, none older than five. The project page video offers a concrete contrast: on one captioning task, image-conditioned forwards drop from 48 to 12, and generation falls from almost 10 seconds to under 3 seconds.

At matched budgets, pruning loses points; LOWN barely does

Per the paper's Table 1, with every method granted at least LOWN's compute budget on LLaDA-V: LOWN retains 99.6% of the baseline score at 0.353x the decoding FLOPs, scoring highest among accelerated methods on eight of ten benchmarks. At the same budget, the visual-token-pruning approaches — RedVTP, ZOO-Prune and KLASS — retain only 87.8%, 88.8% and 89.3% respectively — and KLASS needs per-benchmark thresholds, while LOWN uses one rule throughout. On wall-clock, under H200 at batch size 1, generation time drops 57.4% on average and up to 73.2%. Composition holds up too: stacked with RedVTP at 50% retention it keeps 98.3% at 0.223x FLOPs, highest on all ten benchmarks; stacked with Fast-dLLM's KV cache it cuts time a further 38.1-50.1%. The same rule transfers unchanged to LaViDa — 99.7% retained at 0.174x FLOPs — and extends to VidLaDA under a controlled video protocol.

When the image is needed is a property of the task

The spread from 7.9% to 43.8% says image dependence is task-shaped: document understanding (DocVQA) leans on vision hardest, multiple-choice recognition (MMBench) least. Image-dependent steps also cluster early in the trajectory rather than recurring periodically, so a fixed schedule is doomed; the decision has to be adaptive. This also separates two optimization axes: token pruning controls how much of the image is kept, LOWN controls how often it is consulted, and the two are orthogonal. The paper's example: RedVTP at 50% retention plus LOWN scores 54.9 at 0.195x FLOPs, while RedVTP at 30% retention alone manages 51.7 at 0.354x — cheaper compute, higher score.

Some cold water

First, 99.6% is an average that includes four benchmarks above baseline; capping each at 100% gives 98.6%. Second, stacking with Fast-dLLM is faster but scores lower than Fast-dLLM alone on all three tasks tested, down 5.0 points on LLaVA-Bench — composition is not free. Third, FLOPs are analytical estimates covering only the decoder and LM head, excluding vision encoding, while the timing protocol includes vision encoding; don't mix the two yardsticks. Fourth, every number is author-reported: the GitHub repo has just 5 commits, no third-party replication exists yet, and all backbones are LLaDA-family (LLaDA-V, LaViDa, VidLaDA) — whether this holds on autoregressive VLMs is untested.

For anyone deploying these models, the lesson is direct: visual tokens are the bulk of the multimodal inference bill, and "does this step need the image" is a schedulable axis in its own right, orthogonal to how many tokens the image keeps. Worth flipping over, too: in roughly 77% of denoising steps, looking or not looking unmasks the same token — so which few steps is that re-read image actually feeding?