Research map / Video / Video-language understanding
Keeping the Evidence Chain: Semantic Evidence Allocation for Training-Free Token Pruning in Video Temporal Grounding
Jiaqi Li, Shuntian Zheng, Yixian Shen, JIA-HONG HUANG, Xiaoman Lu, Minzhe Ni, Yu Guan
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 semantic budget allocation to random sampling with random selection (ablation) | SemVID (Qwen3-VL-4B) · Charades-STA 12.5% | mIoU: 41.76 -> 43.26 | Improved |
| add semantic token identification to uniform budget (ablation) | SemVID (Qwen3-VL-4B) · Charades-STA 12.5% | mIoU: 39.80 -> 48.50 | Improved |
| combine budget allocation and token identification (ablation) | SemVID (Qwen3-VL-4B) · Charades-STA 12.5% | mIoU: 48.50 -> 49.89 | Improved |
| plug semantic budget allocation into FastVID (ablation) | FastVID (Qwen3-VL-4B) · Charades-STA 12.5% | mIoU: 35.98 -> 48.88 | Improved |
| use query relevance instead of attention for object selection (ablation) | SemVID object tokens · Charades-STA 12.5% | mIoU: 46.57 -> 49.89 | Improved |
| add MMR diversity to object selection (ablation) | SemVID object tokens · Charades-STA 12.5% | mIoU: 49.44 -> 49.89 | Improved |
| add object tokens (alpha 0 -> 0.6) (ablation) | SemVID · Charades-STA 12.5% | mIoU: 49.48 -> 49.89 | Improved |
| add motion tokens (alpha 1.0 -> 0.6) (ablation) | SemVID · Charades-STA 12.5% | mIoU: 48.11 -> 49.89 | Improved |
| add query-aware filtering of motion tokens (ablation) | SemVID · Charades-STA 12.5% | mIoU: 49.04 -> 49.89 | Improved |
| add 3 context tokens (ablation) | SemVID · Charades-STA 12.5% | mIoU: 49.20 -> 49.89 | Improved |
| increase context tokens from 3 to 10 (ablation) | SemVID · Charades-STA 12.5% | mIoU: 49.89 -> 48.52 | Hurt |
| SemVID vs FastVID on Qwen3-VL-4B under 12.5% token budget | SemVID · ActivityNet-Grounding | mIoU: 33.16 -> 38.49 | 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.
- FastVID — FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025) · better on 7 of 8 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
- DART: Difficulty-Adaptive Routing for Zero-Shot Video Temporal Grounding (ECCV 2026 · similarity 0.91)
- Evidence-Backed Video Question Answering (ECCV 2026 · similarity 0.91)
- Factorized Learning for Temporally Grounded Video-Language Models (ICCV 2025 · similarity 0.91)
- Universal Video Temporal Grounding with Generative Multi-modal Large Language Models (NeurIPS 2025 · similarity 0.90)
- Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding (NeurIPS 2025 · similarity 0.90)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.89)
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.
