VisionWeave routes vision tokens for 2.3× MLLM throughput

VisionWeave, an adaptive visual-encoding method for multimodal LLMs, reports 43.0% average token savings while retaining 98.9% of native benchmark performance across eight tests, and a 2.3× throughput gain in SGLang serving, in an arXiv preprint from Yuan Feng and co-authors posted October 6. By the authors' own comparison, token pruning baselines held only 88% of performance at a fixed 50% savings target.

The paper frames the opening this way: "visual information is unevenly distributed: some regions require fine-grained detail, while others admit compact representations." The system pairs a gated spatial pooler that produces both coarse and fine-grained representations with a granularity router that allocates tokens per region, trained through self-distillation with MRoPE coordinate integration.

Serving metrics in the abstract include a 54.4% mean reduction in time-to-first-token and 60.6% in time-per-output-token. The authors say the approach was validated on Qwen3.5-4B and scaled to Qwen3.8-27B, at a training cost of over 30,000 A100 GPU-hours. Per-benchmark accuracy figures and the hardware configuration behind the SGLang throughput measurement are not in the public summary.