VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs
VisionWeave cuts visual tokens by learning where an MLLM needs detail and where it can stay coarse.
The paper says the method keeps fine and coarse visual representations in a shared coordinate system, then routes granularity by content. On Qwen3.8-27B, it reports 43.0% average token savings while retaining 98.9% native performance across eight benchmarks. In SGLang deployment, the authors report 2.3x throughput, with mean TTFT down 54.4% and mean TPOT down 60.6%. HF Daily Papers' note
The paper says the method keeps fine and coarse visual representations in a shared coordinate system, then routes granularity by content. On Qwen3.8-27B, it reports 43.0% average token savings while retaining 98.9% native performance across eight benchmarks. In SGLang deployment, the authors report 2.3x throughput, with mean TTFT down 54.4% and mean TPOT down 60.6%. HF Daily Papers' note
score 4