Found first: a primary source the press has not covered yet.
A paper posted to arXiv on 6 October 2026 introduces VisionWeave, a method that reduces average visual token counts by 43.0% in multimodal large language models while retaining 98.9% of benchmark performance across eight evaluations, and produces a 2.30× end-to-end throughput gain in SGLang deployment. The paper is available at arXiv:2610.07987.
What the source says
Yuan Feng, Qize Yang, Ruizhe Chen, Sibo Song, and colleagues built VisionWeave on Qwen3.8-27B. The method adds two components: a gated spatial pooler that produces coarse-grained tokens alongside native fine-grained patch tokens within a shared MRoPE coordinate system, and a granularity router that learns content-adaptive token allocation per input. The model is trained via self-distillation across multiple model sizes. On eight benchmarks, VisionWeave retained 98.9% of native performance at 43.0% average token savings. Fixed-50%-pruning baselines retained only 88% performance at the same savings target. SGLang deployment measured a 2.30× end-to-end throughput gain, a 54.4% reduction in mean time to first token, and a 60.6% reduction in mean time per output token.
Why it matters
Fixed-rate visual token pruning forces a static tradeoff: the practitioner picks a compression rate once and accepts the quality loss uniformly. VisionWeave's granularity router allocates tokens per input based on content, so dense or fine-grained scenes get more tokens and simple regions get fewer. The 98.9% performance retention at 43% savings compares favorably against the 88% the paper reports for fixed-50% baselines, which is the more directly comparable configuration. The throughput and latency figures are measured on SGLang, a production inference framework, so teams already running MLLM deployments have a realistic baseline against which to test the method.