Research map / Video / Video-language understanding
TEMPLE: Incentivizing Temporal Understanding of Video Large Language Models via Progressive Pre-SFT Alignment
Shicheng Li, Lei Li, Kun Ouyang, Shuhuai Ren, Yuanxin Liu, Yuanxing Zhang, Fuzheng Zhang, Lingpeng Kong, Qi Liu, Xu Sun
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 Progressive Pre-SFT DPO alignment on self-generated temporal preference data | Qwen2-VL-7B · VideoMME temporal perception | accuracy: 70.9 -> 74.5 | Improved |
| apply Progressive Pre-SFT DPO alignment on self-generated temporal preference data | Qwen2-VL-7B · Vinoground (text) | accuracy: 34.0 -> 37.2 | Improved |
| apply Progressive Pre-SFT DPO alignment on self-generated temporal preference data | Qwen2-VL-7B · Average of four benchmarks | accuracy: 57.7 -> 58.9 vs SFT | Improved |
| apply Progressive Pre-SFT DPO alignment on self-generated temporal preference data | Qwen2-VL-2B · TempCompass (multi-choice) | accuracy: 64.6 -> 64.6 | No clear effect |
| apply DPO before SFT rather than after (ablation) | Qwen2-VL-7B · VideoMME, MLVU, Vinoground | accuracy | Improved |
| use curriculum of progressively harder DPO data (ablation) | Qwen2-VL-7B · Vinoground (text) | accuracy | Improved |
| train on harder DPO data only (single difficulty) (ablation) | Qwen2-VL-7B · Vinoground (text) | accuracy | Hurt |
| apply TEMPLE with DPO data generated by Qwen2-VL-7B (ablation) | Qwen2.5-VL-7B · Vinoground video score | accuracy: 28.8 -> 25.8 | Hurt |
| apply TEMPLE with DPO data generated by Qwen2-VL-7B (ablation) | InternVL2.5-8B · Vinoground text | accuracy: 42.8 -> 45.0 | Mixed |
| use DPO data from Qwen2-VL-7B to tune Qwen2-VL-2B (ablation) | Qwen2-VL-2B · VideoMME, MLVU, Vinoground | accuracy: VideoMME 54.6 -> 52.9, Vinoground text 27.4 -> 26.0 | Hurt |
| use DPO data from Qwen2-VL-2B to tune Qwen2-VL-7B (ablation) | Qwen2-VL-7B · Vinoground text, MLVU | accuracy: Vinoground 34.6 -> 37.4, MLVU 64.6 -> 64.3 | Mixed |
Closest papers by meaning
- Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization (NeurIPS 2025 · similarity 0.93)
- On the Consistency of Video Large Language Models in Temporal Comprehension (CVPR 2025 · similarity 0.90)
- $\mathcal{V}ista\mathcal{DPO}$: Video Hierarchical Spatial-Temporal Direct Preference Optimization for Large Video Models (ICML 2025 · similarity 0.89)
- Momentor: Advancing Video Large Language Model with Fine-Grained Temporal Reasoning (ICML 2024 · similarity 0.89)
- Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025 · similarity 0.88)
- LongVPO: From Anchored Cues to Self-Reasoning for Long-Form Video Preference Optimization (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.
