Research map / Video / Video-language understanding
GTR: Guide-Then-Refine Token Compression for Training-Free Acceleration of Video-LLMs
Sijin Zhou, wei feng, Zhuang Qi, Zhonghua Wang, Lie Ju, Xiang An, Mehrtash Harandi, Zongyuan Ge
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 the global score to the text score (ablation) | GTR (Slocal-free ablation baseline Stext only) · Average relative performance (LLaVA-OV, 25%) | relative performance: 98.3 -> 99.4 | Improved |
| Add the local score on top of text and global scores (ablation) | GTR · Average relative performance (LLaVA-OV, 25%) | relative performance: 99.4 -> 99.7 | Improved |
| Add the local score on top of text and global scores (ablation) | GTR · MLVU | accuracy: 61.7 -> 61.9 | Improved |
| Add the local score on top of text and global scores (ablation) | GTR · LongVideoBench | accuracy: 49.3 -> 49.6 | Improved |
| Use text relevance alone instead of local score alone (ablation) | GTR single-score variants · MLVU | accuracy: 60.0 -> 61.5 | Improved |
| Add GTR to SparseVLM | SparseVLM · MLVU | relative performance: 96.3 -> 97.2 | Improved |
| Add GTR to SparseVLM | SparseVLM · MVBench | relative performance: 99.1 -> 99.6 | Improved |
| Add GTR to FastV | FastV · MLVU | relative performance: 94.6 -> 95.4 | Improved |
| Use GTR token compression at 25% retention instead of full tokens | LLaVA-OV-7B · Total latency | total latency: 26:03 -> 18:44 | Improved |
| Use GTR instead of VidCom2 at 15% retention | LLaVA-OV-7B · Average performance | accuracy: 54.3 -> 55.7 | Improved |
| Reduce retention from 25% to 15% | GTR · Average retained performance (LLaVA-OV) | relative performance: 99.8 -> 95.8 | Hurt |
| Use GTR instead of DyCoke on Qwen2-VL | Qwen2-VL · VideoMME 25% retention | relative performance: 89.5 -> 94.5 | Improved |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.93)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.90)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.90)
- Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective (ICLR 2026 · similarity 0.90)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.90)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (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.
