Xiaomi's 4B Xiaomi-Lens-0 beats Qwen3.5 by 5.6 points on RefCOCO referring tasks
SeerRay-Lab · hf · 2026-10-19
Xiaomi's Xiaoai Plus and SeerRay team released the Xiaomi-Lens-0 technical report: a 4B MLLM built on Qwen3.5 targeting fine-grained region understanding.
- Method: a lightweight decoder fuses vision-encoder patch features with visual latents and the LLM's mask-token states, predicting masks from its own visual representations — no separate segmentation encoder needed.
- Grounded reasoning: each latent decodes to a mask, letting the model interleave chain-of-thought with region-specific latents, effectively "thinking with visual latents" while grounding supervises visited regions.
- Results: gains an average of 5.6 points over Qwen3.5 across all eight RefCOCO/RefCOCO+/RefCOCOg splits, achieves the highest segmentation cIoU among compared models, and surpasses Qwen3.5 on all perception benchmarks.
More from Research
- Six ELLIS members win 2026 AI2050 Fellowships from Schmidt Sciences — maksym_andr · 2026-10-19
- Study of 102 students finds AI use reinforces the illusion of understanding — robertonappe · 2026-10-19
- Anthropic research finds verbalizable representations form a global workspace in LLMs — MoonL88537 · 2026-10-19
- Anthropic researcher shows models pick the rhyme word early, hiding the plan from output — thetripathi58 · 2026-10-19
- Why computational programs likely can't be conscious: Anna Ciaunica's embodied-mind argument — AnnaCiaunica · 2026-10-19
- DNA language models and AlphaGenome answer different questions about variants — Fair-Rain3366 · 2026-10-19