Research map / Video / Video-language understanding
MeToM: Metadata-Guided Token Merging for Efficient Video LLMs
Zhuojie Wu, Shijie Wang, Xin Yu
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 codec-metadata-guided token merging (MeToM) instead of uniform pruning, 50% budget | LLaVA-OneVision-7B · 5 video QA benchmarks average | relative accuracy: 102.0 vs 101.2 (ToMe), 100.4 (HoliTom) | Improved |
| use MeToM at an aggressive 30% token budget | LLaVA-OneVision-7B · 5 video QA benchmarks average | relative accuracy: 101.2 vs 98.4 (FastV), 99.6 (HoliTom) | Improved |
| apply FastV at 30% budget | LLaVA-OneVision-7B · 5 video QA benchmarks average | relative accuracy: 98.4 | Hurt |
| use MeToM on LLaVA-Video-7B at 30% budget | LLaVA-Video-7B · 5 video QA benchmarks average | relative accuracy: 98.4 vs 96.0 (FastV), 97.4 (HoliTom) | Improved |
| use MeToM on Qwen2VL-7B at 30% budget | Qwen2VL-7B · VideoMME, LongVideoBench | relative accuracy: 101.3 vs 97.9 (ToMe) | Improved |
| use MeToM on LLaVA-Video-72B at 50% budget | LLaVA-Video-72B · VideoMME | relative accuracy: 101.4 | Improved |
| use MeToM on LLaVA-Video-72B at 30% budget | LLaVA-Video-72B · VideoMME | relative accuracy: 99.1 vs 97.1 (ToMe), 98.3 (DyCoke*) | Improved |
| add residual-guided patch merging before the vision encoder | MeToM · Vision tower latency | vision tower time: 525ms -> 450ms | Improved |
| use MeToM overall for inference speed | LLaVA-OneVision-7B · TTFT | speedup: 2.65x speedup; LLM backbone 1329ms -> 215ms | Improved |
| use codec-metadata preprocessing instead of similarity-based preprocessing | token compression methods · pre-processing overhead | pre-processing time: 53ms vs HoliTom 92ms, DyCoke 58ms (baseline 50ms) | 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.
- HoliTom — HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025) · better on 3 of 3 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
- HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.91)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.89)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.89)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.89)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.88)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · 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.
