VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs

Key points

• Overcoming these limitations calls for foundation models that learn, end to end, where-and at what granularity-to allocate visual representations, a native capability we term elastic visual representation weaving.1
• We introduce VisionWeave, establishing this capability in frontier-level MLLMs through large-scale training.1
• Based on Qwen3.8-27B, VisionWeave adaptively adjusts token savings to visual content, saving 43.0% tokens on average while retaining 98.9% native performance across eight benchmarks, versus only 88% performance preserved for token pruning baselines with a fixed 50% savings target.1
• Together, we believe these results position elastic visual weaving as a promising capability for next-generation multimodal models.1

Sources (1)

• [1]VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 6, 08:48 AM

  Multimodal large language models have become the dominant paradigm for visual understanding, but incur substantial costs by encoding inputs into dense, fixed-size patch tokens.

  Overcoming these limitations calls for foundation models that learn, end to end, where-and at what granularity-to allocate visual representations, a native capability we term elastic visual representation weaving.

Extractive summary: sentences quoted from the sources.