Research · Oct 30, 2026, 00:00 UTC

Research proposes LOWN, a decoding method that skips image processing in diffusion VLMs

LOWN processes the image only when a no-image prediction disagrees with the last image-conditioned one, retaining 99.6% of LLaDA-V's score on ten benchmarks at 35.3% of decoding FLOPs on average. It needs no training or tuned threshold.

• ModelLLaDA-V
• Topicresearch
• Topicbenchmark