Research map / Video / Video-language understanding
Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs
Jeongseok Hyun, Sukjun Hwang, Su Ho Han, Taeoh Kim, Inwoong Lee, Dongyoon Wee, Joon-Young Lee, Seon Joo Kim, Minho Shim
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 |
|---|---|---|---|
| use STTM instead of other query-agnostic merging (ToMe, DyCoke-stage1) | LLaVA-Video-7B, 30% token budget · VNBench | relative accuracy: 82.7 -> 98.0 (vs ToMe) | Improved |
| use multi-granular quadtree spatial merging instead of single-granularity merging (ablation) | STTM spatial merging · VNBench | relative accuracy: 80.0 -> 98.7 | Improved |
| single-granularity merging on VideoMME (ablation) | STTM spatial merging · VideoMME | relative accuracy: 98.4 vs 99.2 | No clear effect |
| use 4x4 root nodes instead of 2x2 in quadtree (ablation) | STTM spatial merging · VNBench | relative accuracy: 73.1 -> 98.7 | Improved |
| add temporal merging after spatial merging (sequential decomposition) (ablation) | STTM · VNBench / VideoMME | relative token count: R.NV 57.2 -> 25.8 on VNBench with accuracy 98.7 -> 98.0 | Improved |
| joint octree spatio-temporal merging instead of decomposed merging (ablation) | STTM · VNBench | relative accuracy: 98.0 -> 73.3 | Hurt |
| use top-left approximate merge destination instead of most similar (ablation) | STTM temporal merging · 30% budget average | relative accuracy: 97.7 -> 98.8; TTFT 39.6 -> 31.4 | No clear effect |
| reassign position ids instead of merging or keeping RoPE (ablation) | STTM positional embedding · VNBench | relative accuracy: 96.0 -> 98.0 | Improved |
| merge tokens at later LLM layer (19 instead of 3) (ablation) | STTM · VNBench | relative accuracy vs TTFT: accuracy 98.0 -> 99.5; relative TTFT 30.9 -> 75.5 | Mixed |
| remove special grid token in LLaVA-Video (ablation) | LLaVA-Video-7B · 6 benchmarks average | accuracy: 68.5 -> 68.8 | No clear effect |
| apply STTM to larger or different MLLMs (OneVision, Qwen2VL, 72B) | STTM · OneVision/Qwen2VL/LLaVA-Video-72B | relative accuracy: 102.1 (OneVision 50%), 102.7 (Qwen2VL 50%), 101.3 (72B 50%) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FrameFusion — FrameFusion: Combining Similarity and Importance for Video Token Reduction on Large Vision Language Models (ICCV 2025) · better on 1 of 1 shared result
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
- HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.91)
- Granularity-Adaptive Spatial Evidence Tokenization for Video Question Answering (AAAI 2025 · similarity 0.89)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.89)
- Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025 · similarity 0.88)
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.88)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.88)
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.
