arXiv 2610.23279
Xiaomi-Lens-0 Technical Report
AI12 upvotes
Authors: Pei Fu, Yike Liu, Yue Sun, Hanmo Wang, Xiuwen Xi, Ruoceng Zhang, Shaojie Zhang, Chao Jiang +7
Overview
The authors present Xiaomi-Lens-0, a 4B multimodal LLM built on Qwen3.5 that predicts segmentation masks directly from its own visual representations. A lightweight decoder fuses vision-encoder patch features with the language model's mask-token states, so mask supervision reaches the shared backbone without a separate segmentation encoder. Because each visual latent decodes to a mask, the model can interleave region-specific latents into its reasoning and supervise the regions it visits during chain-of-thought.
Introduction
Multimodal LLMs answer whole-image questions well but struggle with fine-grained region understanding. Text-only supervision never specifies which regions support an answer, which is a key reason for this gap.
Results
Xiaomi-Lens-0 gains 5.6 points on average over Qwen3.5 across all eight RefCOCO, RefCOCO+, and RefCOCOg splits in referring comprehension. It also achieves the highest segmentation cIoU among compared models and surpasses Qwen3.5 on all perception benchmarks.
Conclusion
Grounding reasoning in predicted masks offers a practical route to better fine-grained region understanding in MLLMs.