Research map / Video / Video-language understanding
Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention
Junhao Du, Jialong Xue, Anqi Li, Jincheng Dai, Guo Lu
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 similarity-based pruning to attention selection (ablation) | attention-only selection · LLaVA-OV-7B 2% retention, 5 benchmarks | Avg. score: 41.8 -> 49.7 | Improved |
| add clustering-based merging (ablation) | attention + similarity · LLaVA-OV-7B 2% retention, 5 benchmarks | Avg. score: 49.7 -> 50.4 | Improved |
| add attention scores (ablation) | similarity + cluster · LLaVA-OV-7B 2% retention, 5 benchmarks | Avg. score: 49.5 -> 50.4 | Improved |
| add similarity pruning (ablation) | attention + cluster · LLaVA-OV-7B 2% retention, 5 benchmarks | Avg. score: 45.8 -> 50.4 | Improved |
| add text-aware inner-LLM merging (ablation) | unified spatiotemporal compression · LLaVA-OV-7B 2% retention | Avg. % of baseline: 89.5 -> 90.1 | Improved |
| add text-aware inner-LLM merging (ablation) | unified spatiotemporal compression · LLaVA-OV-7B 5% retention | Avg. % of baseline: 94.9 -> 95.4 | Improved |
| add text-aware inner-LLM merging (ablation) | unified spatiotemporal compression · LLaVA-OV-7B 1% retention | Avg. % of baseline: 84.2 -> 84.1 | No clear effect |
| use text-aware scoring instead of last-token attention (ablation) | inner-LLM pruning · LLaVA-OV-7B 5% retention | Avg. % of baseline: 93.8 -> 95.4 | Improved |
| raise similarity threshold tau from 0.7 to 0.9 (ablation) | Ours w/o M · LLaVA-OV-7B 2% retention, 5 benchmarks | Avg. %: 89.5 -> 86.0 | Hurt |
| prune inside LLM at layer 18 instead of layer 2 (ablation) | Ours · LLaVA-OV-7B 5 benchmarks | Avg. %: 83.6 -> 90.1 | Improved |
| set cluster ratio 0.3 instead of 0.7 (ablation) | Ours w/o M · LLaVA-OV-7B 2% retention, 5 benchmarks | Avg. %: 87.6 -> 89.5 | Improved |
| set lambda 0.5 instead of 0 in text-aware merging (ablation) | Ours · LLaVA-OV-7B 5 benchmarks | Avg. %: 89.7 -> 90.1 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- HoliTom — HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025) · better on 9 of 9 shared results
- FastVID — FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.91)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.91)
- Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective (ICLR 2026 · similarity 0.90)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.90)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.90)
- Recurrent Attention-based Token Selection for Efficient Streaming Video-LLMs (NeurIPS 2025 · 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.
