Xiaomi‑Lens‑0 is a 4B multimodal large language model built on Qwen3.5 that predicts region masks directly from its visual representations, eliminating the need for a separate segmentation encoder. By combining vision‑encoder patch features with visual latents and mask‑token states, the model interleaves reasoning with region‑specific latents, enabling grounded chain‑of‑thought reasoning. Across all RefCOCO, RefCOCO+, and RefCOCOg splits, Xiaomi‑Lens‑0 outperforms Qwen3.5 by an average of 5.6 points in referring comprehension and achieves the highest segmentation cIoU among compared models, while also surpassing Qwen3.5 on all perception benchmarks.
arXiv:2502.04192v5 Announce Type: replace Abstract: Multiple works have emerged to push the boundaries of multi-modal large language models (MLLMs) towards pixel-level understanding. The current tren...
arXiv:2610.32664v1 Announce Type: new Abstract: Multimodal large language models (MLLMs) read images through visual tokens, and under the small per-image budgets that efficient deployment requires, t...
arXiv:2608.28707v1 Announce Type: new Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable progress in Visual Question Answering (VQA), yet they continue to struggle with quest...
arXiv:2605.26380v2 Announce Type: replace-cross Abstract: Frontier multimodal large language models (MLLMs) have been reported to achieve over 90\% accuracy on fine-grained perception benchmarks. How...
Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces. We ask whether CoT requires continued access to image tokens, or whether it mainly operates over visual information already made available earlier in the forward pass.
The paper introduces MIMO-Bench, a new benchmark designed to test Multimodal Large Language Models (MLLMs) on visual reasoning tasks that involve multiple instances of the same object category. Unlike existing benchmarks that focus on images with few object instances, MIMO-Bench contains 3,802 image-question pairs requiring the model to disambiguate among similar objects using attribute binding and numerical counting. Results show that state‑of‑the‑art models, both open‑source and proprietary, perform significantly worse on this benchmark, underscoring their limitations in handling multi‑instance, multi‑object scenarios.
arXiv:2607. 12815v1 Announce Type: new Abstract: Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces.
MultiInpaintCOCO extends the English InpaintCOCO dataset to 13 typologically diverse languages, creating a multilingual benchmark of 953 minimal-pair examples focused on fine-grained visual details. The study evaluates state-of-the-art multilingual multimodal retrieval models, revealing that performance drops significantly for lower-resource languages and that many models exhibit a systematic bias toward English captions. The benchmark and code are publicly available at https://github.com/ufal/multi-inpaint-coco.
The paper introduces Visual Retrieval Heads (VRHs), a small fraction of attention heads in vision‑language models that are causally responsible for grounding text descriptions to image regions. By recasting head‑scoring methods and evaluating across eleven VLMs and five benchmarks, the authors show that masking the top 20 VRHs can drop grounding accuracy by up to 80 percentage points, while random masking has little effect. VRHs generalize across various visual reference tasks, preserve output format while corrupting localization, and transfer causally across models sharing an LLM backbone.
MultiInpaintCOCO extends the English InpaintCOCO dataset to 13 typologically diverse languages, creating a multilingual benchmark of 953 minimal-pair examples focused on fine-grained visual details. The study evaluates state-of-the-art multilingual multimodal retrieval models, revealing that performance drops significantly for lower-resource languages and that many models exhibit a systematic bias toward English captions. The benchmark and code are publicly available at https://github.com/ufal/multi-inpaint-coco.
The paper introduces SAGE, a lightweight method that guides decoder attention in vision-language models away from prompt‑invariant attention sinks (PIS) toward query‑dependent regions of interest (ROIs). By using token‑aligned ROI masks from vision backbones like CLIP, ViT, and DINOv3, SAGE improves visual grounding and reduces hallucinations across various VLM families. Experiments show consistent gains on public vision‑language benchmarks, especially in fine‑grained visual discrimination tasks where localized evidence is essential.
LensVLM is an inference framework and post‑training recipe that lets Vision‑Language Models (VLMs) process compressed images of text by selectively expanding only the relevant parts back to full resolution. Using Qwen3.5‑9B‑Base, LensVLM achieves accuracy comparable to full‑text models at 4.3× compression and outperforms other compression baselines up to 10.1× across seven text QA benchmarks, while also improving performance on multimodal document and code tasks as compression increases.