Research map / Video / Video-language understanding
MARC: Memory-Augmented RL Token Compression for Efficient Video Understanding
Peiran Wu, Zhuorui Yu, Yunze Liu, Chi-Hao Wu, Enmin Zhou, Junxiao Shen
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 Visual Memory Retriever (top-3 clips) to Qwen2.5-VL-3B 64 frames (ablation) | Qwen2.5-VL-3B · 6 video benchmarks (mean) | accuracy: 42.21 -> 45.56 | Improved |
| add VMR to Qwen2.5-VL-3B on VideoMME (ablation) | Qwen2.5-VL-3B · VideoMME (w/o sub) | accuracy: 53.55 -> 51.85 | Hurt |
| add VMR to Qwen2.5-VL-3B on VideoMMMU (ablation) | Qwen2.5-VL-3B · VideoMMMU | accuracy: 35.33 -> 34.33 | Hurt |
| use C-GRPO instead of SFT for 1-frame compression (ablation) | MARC-3B · 6 video benchmarks (mean) | accuracy: 38.50 -> 42.20 | Improved |
| use C-GRPO instead of SFT, both with VMR (ablation) | MARC-3B · 6 video benchmarks (mean) | accuracy: 40.16 -> 42.20 | Improved |
| add VMR to SFT-trained 1-frame model (ablation) | Qwen2.5-VL-3B SFT · 6 video benchmarks (mean) | accuracy: 38.50 -> 40.16 | Improved |
| set retention threshold tau 0.6 instead of 0.4 (ablation) | MARC-3B · 6 video benchmarks (mean) | accuracy: 41.34 -> 42.20 | Improved |
| set retention threshold tau 0.6 instead of 0.8 (ablation) | MARC-3B · 6 video benchmarks (mean) | accuracy: 41.40 -> 42.20 | Improved |
| compress to 1 frame-equivalent tokens with MARC vs 64-frame baseline | MARC-3B · VideoMME (w/o sub) | accuracy: 53.55 -> 39.44 | Hurt |
| compress to 1 frame-equivalent tokens with MARC vs 64-frame baseline | MARC-3B · VSI-Bench | accuracy: 32.93 -> 27.55 | Hurt |
| compress to 1 frame-equivalent tokens with MARC vs 64-frame baseline | MARC-3B · TempCompass | accuracy: 38.05 -> 55.34 | Improved |
| compress to 1 frame-equivalent tokens with MARC vs 64-frame baseline | MARC-3B · 6 video benchmarks (mean) | accuracy: 42.21 -> 42.20 | No clear effect |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.92)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.91)
- Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective (ICLR 2026 · similarity 0.91)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.91)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.90)
- EvoComp: Learning Visual Token Compression for Multimodal Large Language Models via Semantic-Guided Evolutionary Labeling (CVPR 2026 · similarity 0.90)
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.
