Research map / Video / Video-language understanding
PAS: A Training-Free Stabilizer for Temporal Encoding in Video LLMs
Bowen Sun, Yujun Cai, Ming-Hsuan Yang, Hang Wu, Yiwei 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 |
|---|---|---|---|
| apply PAS (phase offsets [0, 0.5] across two head groups) to the backbone | Qwen2.5-VL-7B-Instruct · Kinetics-700 | accuracy: 44.9 -> 48.2 | Improved |
| apply PAS to the backbone | Qwen2.5-VL-7B-Instruct · 20BN-Jester | accuracy: 16.0 -> 18.3 | Improved |
| apply PAS to the backbone | Qwen2.5-VL-7B-Instruct · MVBench | accuracy: 67.2 -> 69.5 | Improved |
| stack PAS on SlowFast-LLaVA | SlowFast-LLaVA · Kinetics-700 | accuracy: 45.1 -> 49.8 | Improved |
| stack PAS on TS-LLaVA | TS-LLaVA · MMBench-Video | mean score: 1.70 -> 1.76 | No clear effect |
| apply PAS on sub-Nyquist sampled Breakfast Actions | Qwen2.5-VL-7B-Instruct · Breakfast Actions | accuracy: 44.5 -> 45.3 | No clear effect |
| vary phase offset magnitude Delta from 0.0 to 1.0 (K=2) (ablation) | PAS · Jester, SSv2, Kinetics-700 | accuracy: gains for Delta in 0.3 to 0.8; default 0.5 | Improved |
| offset magnitude near 1.0 (bin-scale aliasing boundary) (ablation) | PAS · action benchmarks | accuracy | No clear effect |
| increase sampling ratio r (denser frame sampling) (ablation) | PAS · Jester, SSv2, Kinetics-700 | accuracy gain over backbone | Hurt |
| enable PAS at inference | Qwen2.5-VL-7B-Instruct · A100 throughput | tokens/s: (77.2±3.1)×103 -> (76.8 ± 4.0) × 103 | No clear effect |
| stack PAS on SlowFast-LLaVA | SlowFast-LLaVA · MVBench | accuracy: 69.2 -> 71.0 | Improved |
Closest papers by meaning
- A Simple Recipe for Contrastively Pre-training Video-First Encoders Beyond 16 Frames (CVPR 2024 · similarity 0.86)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.86)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.85)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.85)
- TC-LLaVA: Rethinking the Transfer of LLava from Image to Video Understanding with Temporal Considerations (AAAI 2025 · similarity 0.85)
- EchoPrune: Interpreting Redundancy as Temporal Echoes for Efficient VideoLLMs (NeurIPS 2026 · similarity 0.85)
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.
