Research map / Video / Video-language understanding
S3-Prune: Stability-Aware Token Budgeting for Long-Form Video-Language Models
gilha lee, Seungil Lee, Hyun Kim
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 Kalman filter stability accumulation (ablation) | S3-prune on LLaVA-OV · 4 benchmarks (10% tokens) | avg score: 53.6 -> 54.4 | Improved |
| add Kalman filter stability accumulation (ablation) | S3-prune on LLaVA-Video · 4 benchmarks (10% tokens) | avg score: 57.3 -> 58.1 | Improved |
| add Kalman filter stability accumulation (ablation) | S3-prune on Qwen2.5-VL-7B · 4 benchmarks (10% tokens) | avg score: 59.6 -> 60.2 | Improved |
| add spatial uncertainty signal to segment transition signal (ablation) | S3-prune Qwen3-VL-8B · MVBench+VideoMME | avg score: 59.7 -> 60.2 | Improved |
| add segment transition signal to spatial uncertainty signal (ablation) | S3-prune Qwen3-VL-8B · MVBench+VideoMME | avg score: 59.9 -> 60.2 | Improved |
| add S3-prune budgeting to FastVID segmentation | FastVID · 4 benchmarks (10% tokens) | avg score: 59.3 -> 59.7 | Improved |
| add S3-prune budgeting to HoliTom segmentation | HoliTom · 4 benchmarks (10% tokens) | avg score: 57.2 -> 57.7 | Improved |
| use S3-prune pre-LLM stage instead of FastVid | FastVid · MVBench+VideoMME avg (15% tokens) | avg score: 61.4 -> 61.9 | Improved |
| use S3-prune instead of HoliTom at 10% retention | HoliTom · LLaVA-OV-7B 7 benchmarks | avg score: 59.6 -> 59.7 | Improved |
| use S3-prune at 25% token retention | LLaVA-OV-7B · 7 benchmarks | avg score: 60.7 -> 61.1 | Improved |
| use S3-prune at 25% token retention | LLaVA-Video-7B · 7 benchmarks | avg score: 65.9 -> 65.8 | No clear effect |
| use S3-prune at 10% retention instead of HoliTom | HoliTom · Qwen3-VL-8B 4 benchmarks | avg performance retained (%): 88.1 -> 93.3 | Improved |
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.
- FastVid — FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025) · better on 8 of 8 shared results
- HoliTom — HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025) · better on 6 of 6 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
- Token Reduction via Local and Global Contexts Optimization for Efficient Video Large Language Models (CVPR 2026 · similarity 0.89)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.88)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.87)
- HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.87)
- ST3: Accelerating Multimodal Large Language Model by Spatial-Temporal Visual Token Trimming (AAAI 2025 · similarity 0.86)
- Sharp Eyes and Memory for VideoLLMs: Information-Aware Visual Token Pruning for Efficient and Reliable VideoLLM Reasoning (AAAI 2026 · similarity 0.86)
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.
