Research map / Video / Video-language understanding
EarlyTom: Early Token Compression Completes Fast Video Understanding
Hesong Wang, Xin Jin, Lu Lu, Chenhaowen Li, Jian Chen, Qiang Liu, Huan Wang
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 decoupled spatial token selection to frame merging (ablation) | EarlyTom, LLaVA-OV-7B · 3 video benchmarks avg, 20% retained | avg score: 58.4 -> 58.8 | Improved |
| add inner-encoder frame merge to spatial selection (ablation) | EarlyTom, LLaVA-OV-7B · 3 video benchmarks avg, 20% retained | avg score: 58.4 -> 58.8 | Improved |
| local window sampling instead of random sampling (ablation) | EarlyTom, LLaVA-OV-7B · 3 video benchmarks avg, 20% retained | avg score: 57.8 -> 58.8 | Improved |
| local window sampling instead of global top-K (ablation) | EarlyTom, LLaVA-OV-7B · 3 video benchmarks avg, 20% retained | avg score: 58.4 -> 58.8 | Improved |
| local window sampling instead of global top-K (ablation) | EarlyTom, LLaVA-OV-7B · throughput (tokens/s) | throughput: 31.5 -> 33.4 | Improved |
| local window sampling instead of random sampling (ablation) | EarlyTom, LLaVA-OV-7B · throughput (tokens/s) | throughput: 35.3 -> 33.4 | Hurt |
| start frame merging at layer 6 instead of layer 4 (ablation) | EarlyTom, LLaVA-OV-7B · avg score, 20% retained | avg score: 58.6 -> 58.9 | Improved |
| start frame merging at layer 10 instead of layer 6 (ablation) | EarlyTom, LLaVA-OV-7B · TTFT (ms), 20% retained | TTFT: 387.1 -> 436.9 | Hurt |
| add decoupled spatial token selection (ablation) | EarlyTom, Qwen2.5-VL-7B · avg score, 15% retained | avg score: 61.4 -> 62.2 | Improved |
| add weighted frame merging (ablation) | EarlyTom, Qwen2.5-VL-7B · avg score, 15% retained | avg score: 61.3 -> 62.2 | Improved |
| compress inside vision encoder vs HoliTom late compression | EarlyTom, LLaVA-OV-7B · TTFT (ms), 10% retained | TTFT: 556.6 -> 336.2 | Improved |
| EarlyTom vs HoliTom accuracy | EarlyTom, LLaVA-OV-7B · 4-benchmark avg, 10% retained | avg score: 57.9 -> 56.2 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- HoliTom — HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025) · better on 4 of 8 shared results
- FastVID — FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025) · better on 3 of 4 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)
- PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.91)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.91)
- DUET-VLM: Dual stage Unified Efficient Token reduction for VLM Training and Inference (CVPR 2026 · similarity 0.90)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.90)
- HiRED: Attention-Guided Token Dropping for Efficient Inference of High-Resolution Vision-Language Models (AAAI 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.
