Research map / Vision-language and multimodal / Multimodal LLMs
Multiple Images Distract Large Multimodal Models via Attention Fragmentation
Tingrui Qiao, Di Zhao, Yuzhuo Li, Bo Pang, Caroline Walker, Chris Cunningham, Yun Sing Koh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Attention Remasking (AR) | Qwen3-VL-8B · MMIU | accuracy (%): 55.6 -> 59.4 | Improved |
| add Attention Remasking (AR) | Qwen3-VL-8B · MuirBench | accuracy (%): 51.3 -> 54.9 | Improved |
| add Attention Remasking (AR) | LLaVA-OneVision-7B · MMIU | accuracy (%): 36.9 -> 41.2 | Improved |
| add Attention Remasking (AR) | InternVL3.5-8B · MMIU | accuracy (%): 49.3 -> 54.1 | Improved |
| add Attention Remasking (AR) | DeepSeek-VL2-Small · MIRB | accuracy (%): 45.8 -> 50.7 | Improved |
| add Attention Remasking (AR) | Qwen3-VL-32B · MMIU | accuracy (%): 59.8 -> 62.6 | Improved |
| add Attention Remasking (AR) | DeepSeek-VL2-27B · MuirBench | accuracy (%): 43.7 -> 46.8 | Improved |
| add VAR post-softmax redistribution | Qwen3-VL-8B · MMIU | accuracy (%): 55.6 -> 55.9 | No clear effect |
| add SoFA attention interpolation | Qwen3-VL-8B · MMIU | accuracy (%): 55.6 -> 56.7 | Improved |
| take grounding prior from last layer instead of layer 1 (ablation) | AR on Qwen3-VL-8B · MMIU | accuracy (%): 59.4 -> 53.2 | Hurt |
| take grounding prior from last layer instead of layer 1 (ablation) | AR on InternVL3.5-8B · MMIU | accuracy (%): 54.1 -> 46.5 | Hurt |
| take grounding prior from middle layer instead of layer 1 (ablation) | AR on Qwen3-VL-8B · MMIU | accuracy (%): 59.4 -> 54.8 | Hurt |
Closest papers by meaning
- See What You Are Told: Visual Attention Sink in Large Multimodal Models (ICLR 2025 · similarity 0.92)
- Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models (ICML 2025 · similarity 0.87)
- Attention (as Discrete-Time Markov) Chains (NeurIPS 2025 · similarity 0.87)
- Zooming from Context to Cue: Hierarchical Preference Optimization for Multi-Image MLLMs (NeurIPS 2025 · similarity 0.87)
- From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning (ICLR 2026 · similarity 0.87)
- Blink: Dynamic Visual Token Resolution for Enhanced Multimodal Understanding (CVPR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
