Research map / Video / Video-language understanding
MSJoE: Jointly Evolving MLLM and Sampler for Efficient Long-Form Video Understanding
Wenhui Tan, Xiaoyi Yu, Jiaze Li, Yijing Chen, Jianzhong Ju, Zhenbo Luo, Ruihua Song, Jian Luan
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 |
|---|---|---|---|
| jointly train MLLM and sampler instead of freezing the MLLM (ablation) | MSJoE · VideoMME-Long (32 frames) | accuracy: 52.2 -> 54.1 | Improved |
| train the MLLM to generate queries (vs frozen MLLM) with same sampled frames (ablation) | MSJoE · LVBench (32 frames) | accuracy: 44.8 -> 46.4 | Improved |
| add sampler pre-training before joint RL (ablation) | MSJoE · MLVU, LoVi, VLong, LVB (32 frames) | accuracy: MLVU 69.1 -> 69.3, LVBench 46.2 -> 46.4 | Improved |
| train the sampler during RL instead of keeping it frozen (MLLM trained) (ablation) | MSJoE · LongVideoBench (32 frames) | accuracy: 58.1 -> 60.1 | Improved |
| use the question as CLIP query with top-k instead of uniform sampling (ablation) | Qwen2.5-VL-7B frozen · MLVU, VideoMME-Long, LVBench (32 frames) | accuracy: MLVU 61.5 -> 68.3, VLong 49.9 -> 48.9 (lower) | Improved |
| decompose question into multiple queries with top-k pooling (no trained sampler) (ablation) | frozen Qwen2.5-VL-7B · MLVU (32 frames) | accuracy: 68.3 -> 64.5 | Hurt |
| replace top-k with a trained sampler for multi-query similarity (frozen MLLM) (ablation) | frozen Qwen2.5-VL-7B · VideoMME-Long (32 frames) | accuracy: 45.4 -> 52.2 | Improved |
| naive top-k similarity sampling vs uniform sampling at varying frame budgets (ablation) | Qwen2.5-VL-7B · LongVideoBench and VideoMME-Long | accuracy | Hurt |
| apply MSJoE instead of uniform sampling | Qwen2.5-VL-7B (64 frames) · LVBench | accuracy: 39.2 -> 51.1 | Improved |
| apply MSJoE instead of TSPO (learned sampler with frozen MLLM) | Qwen2.5-VL-7B (64 frames) · MLVU, VideoMME-Long, LVBench | accuracy: MLVU 74.3 -> 75.1, LoVi 64.2 -> 62.2 | Mixed |
Closest papers by meaning
- Adaptive Keyframe Sampling for Long Video Understanding (CVPR 2025 · similarity 0.91)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.90)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.90)
- Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025 · similarity 0.90)
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.89)
- SVBench: A Benchmark with Temporal Multi-Turn Dialogues for Streaming Video Understanding (ICLR 2025 · 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.
