Xiaomi-Lens-0 Technical Report
Pei Fu, Yike Liu, Yue Sun, Hanmo Wang, Xiuwen Xi, Ruoceng Zhang, Shaojie Zhang, Tingwei Huang et al. · SeerRay-Lab
16 Oct 2026 · cs.CV · tech report · single gpu · code · reasoning & test-time compute, multimodal, evaluation & benchmarks
52 hf upvotes · 0 hn points, 0 comments · 105 stars · 0 citations
Multimodal large language models (MLLMs) answer questions about whole images well, yet they still struggle with fine-grained region understanding. A key reason is that text-only supervision never specifies which regions support an answer. To close this gap, we present Xiaomi-Lens-0, a 4B MLLM built on Qwen3.5 that predicts masks from its own visual representations. Specifically, a lightweight decoder combines vision-encoder patch features with the language model's mask-token states, termed visual latents, so that mask supervision reaches the shared backbone without an additional segmentation encoder. Because each latent decodes to a mask, the model can interleave reasoning with region-specific latents and thus think with visual latents, while grounded reasoning traces supervise the intermediate regions referenced during chain-of-thought. As a result, across all eight RefCOCO, RefCOCO+, and RefCOCOg splits, Xiaomi-Lens-0 gains 5.6 points on average over Qwen3.5 in referring comprehension and achieves the highest segmentation cIoU among compared models. Moreover, it surpasses Qwen3.5 across all perception benchmarks. Homepage: https://seerray-lab.github.io/Xiaomi-Lens-0
Compute guess: single gpu. Based on: smallest model mentioned ~4B params; small-scale or toy setting. (estimated from the abstract)
Potential
signals behind the hidden gems rankingemerging topic “Multimodal Large Language Models”: 16 papers in 2 weeks vs 19 in the 18 days before · shares 3 authors with “A-RSI: Adaptive Task Refresh for Zero-Data Multimodal Self-I…” (75 upvotes/points)
Existing reproductions
Not checked yet. Looks for GitHub repos that cite this arXiv id and HF models tagged with it, then separates independent reproductions from code that only cites or uses the paper, and from paper lists.
Comments