Research map / Video / Video-language understanding
SURGE: Surprise-Guided Token Reduction for Efficient Video Understanding with VLMs
Chong Tang, Sannara Ek, Dirk Koch, Robert D. Mullins, Alex S. Weddell, Jagmohan Chauhan
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 global affine drift detrending (ablation) | SURGE · MLVU M-Avg / G-Avg, Qwen2.5-VL rho=0.25 | M-Avg / G-Avg: 64.9 / 4.18 -> 65.7 / 4.26 | Improved |
| add variance normalization (ablation) | SURGE · MLVU M-Avg / G-Avg, Qwen2.5-VL rho=0.25 | M-Avg / G-Avg: 65.1 / 4.22 -> 65.7 / 4.26 | Improved |
| add causal constant-velocity temporal predictor over frame differences (ablation) | SURGE · TempCompass, Qwen2.5-VL rho=0.25 | accuracy: 66.9 -> 70.5 | Improved |
| add CLIP Top-5 event focusing to SURGE (ablation) | SURGE on Qwen2.5-VL · Video-MME | accuracy: 60.9 -> 62.0 | Improved |
| use K=1 instead of K=5 in CLIP event focusing (ablation) | SURGE* · Video-MME | accuracy: 62.0 -> 51.7 | Hurt |
| replace SURGE with random token pruning at rho=0.01 (ablation) | Qwen2.5-VL · Video-MME | accuracy: 58.7 -> 29.8 | Hurt |
| replace SURGE with random token pruning at rho=0.25 (ablation) | Qwen2.5-VL · Video-MME | accuracy: 60.9 -> 53.6 | Hurt |
| use SURGE instead of FastV pruning | Video-LLaVA-Qwen (7B) · Video-MME | accuracy: 58.1 -> 64.5 | Improved |
| combine AKS keyframe selection with SURGE | AKS on Video-LLaVA-Qwen · Video-MME | accuracy: 64.7 -> 64.9 | Improved |
| apply SURGE at rho=0.25 to full tokens | Qwen2.5-VL (7B) · Video-MME | accuracy: 62.2 -> 60.9 | Hurt |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs (NeurIPS 2026 · similarity 0.89)
- FlexSelect: Flexible Token Selection for Efficient Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.87)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.87)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · 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.
