Research map / Video / Video-language understanding
Free-MoRef: Instantly Multiplexing Context Perception Capabilities of Video-MLLMs within Single Inference
Kuo Wang, Quanlong Zheng, Junlin Xie, Yanhao Zhang, Jinguo Luo, Haonan Lu, Liang Lin, Fan Zhou, Guanbin Li
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 |
|---|---|---|---|
| apply Free-MoRef to 128-frame input | LLaVA-Video-7B · VideoMME overall | accuracy: 64.9 -> 66.3 | Improved |
| apply Free-MoRef to 256-frame input | LLaVA-Video-7B · MLVU | accuracy: 67.2 -> 72.5 | Improved |
| apply Free-MoRef to 512-frame input | LLaVA-Video-7B · VideoMME overall | accuracy: 60.6 -> 66.9 | Improved |
| apply Free-MoRef to 512-frame input | LLaVA-Video-7B · LongVideoBench overall | accuracy: 54.3 -> 59.9 | Improved |
| double input frames from 64 to 128 without Free-MoRef | LLaVA-Video-7B · MLVU | accuracy: 70.3 -> 70.2 | No clear effect |
| quadruple input frames from 64 to 256 without Free-MoRef | LLaVA-Video-7B · VideoMME overall | accuracy: 64.3 -> 63.1 | Hurt |
| add MoRef attention to partition plus fusion (ablation) | Free-MoRef · VideoMME overall, 128 frames | accuracy: 62.0 -> 66.3 | Improved |
| add Reference Fusion to partition plus MoRef attention (ablation) | Free-MoRef · VideoMME overall, 128 frames | accuracy: 65.8 -> 66.3 | Improved |
| add Reference Fusion alone (drop 50% tokens, FastV-like) without partition (ablation) | LLaVA-Video-7B · VideoMME overall, 128 frames | accuracy: 64.9 -> 63.9 | Hurt |
| add multi-reference partition with full attention instead of MoRef attention (ablation) | Reference Fusion baseline · VideoMME overall, 128 frames | accuracy: 63.9 -> 62.0 | Hurt |
| increase chunk number N from 2 to 8 (ablation) | Free-MoRef · VideoMME overall, 128 frames | accuracy: 66.3 -> 65.9 | Hurt |
| change temporal units M from 1 to 64 (ablation) | Free-MoRef · VideoMME temporal perception | accuracy: 70.9 -> 74.5 | Improved |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- VideoICL: Confidence-based Iterative In-context Learning for Out-of-Distribution Video Understanding (CVPR 2025 · similarity 0.89)
- POINTS-Long: Adaptive Dual-Mode Visual Reasoning in MLLMs (CVPR 2026 · similarity 0.89)
- Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.89)
- Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification (ICLR 2025 · similarity 0.89)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.89)
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.
