Research map / Video / Video-language understanding
Thinking with Drafts: Speculative Temporal Reasoning for Efficient Long Video Understanding
Pengfei Hu, Meng Cao, Yingyao Wang, Yi Wang, Jiahua Dong, Jun Song, Yu Cheng, Bo Zheng, Xiaodan Liang
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 |
|---|---|---|---|
| replace uniform 16-frame sampling with target+draft frame selection (ablation) | SpecTemp · Video-Holmes | accuracy: 39.5 -> 47.0 | Improved |
| add target-predicted temporal region before uniform sampling (ablation) | Uniform sampling · Video-Holmes | accuracy: 39.5 -> 43.2 | Improved |
| use CLIP similarity for frame selection after region prediction (ablation) | Target + Uniform · Video-Holmes | accuracy: 43.2 -> 45.1 | Improved |
| replace CLIP with a trained 3B draft model for frame selection (ablation) | Target + CLIP · Video-MME (wo sub) | accuracy: 60.5 -> 62.4 | Improved |
| use 7B-only for both roles instead of 7B target + 3B draft (ablation) | SpecTemp · LongVideoBench | accuracy and latency: 57.5 -> 54.1 acc; 2.3s -> 3.1s | Hurt |
| use 3B-only for both roles instead of 7B target + 3B draft (ablation) | SpecTemp · LongVideoBench | accuracy: 57.5 -> 40.3 | Hurt |
| add SFT to target model (ablation) | Qwen2.5-VL-7B target · LongVideoBench | accuracy: 54.5 -> 55.3 | Improved |
| add RL to target model after SFT (ablation) | Target model · LongVideoBench | accuracy: 55.3 -> 56.3 | Improved |
| add draft model SFT (ablation) | Target SFT · LongVideoBench | format accuracy: 95.2 -> 98.7 | Improved |
| add RL to draft model (full SFT+RL for both) (ablation) | SpecTemp · LongVideoBench | accuracy: 57.1 -> 57.5 | Improved |
| use SpecTemp at about 14 frames instead of Qwen2.5-VL-7B 16 frames | Qwen2.5-VL-7B · Video-Holmes | accuracy: 35.0 -> 47.0 | Improved |
| use SpecTemp instead of VideoChat-R1.5 at 64-frame scale | VideoChat-R1.5 · long-form latency | latency (s): 11.5 -> 8.9 | Improved |
Closest papers by meaning
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.90)
- Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs (NeurIPS 2026 · similarity 0.90)
- TOMATO: Assessing Visual Temporal Reasoning Capabilities in Multimodal Foundation Models (ICLR 2025 · similarity 0.89)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.89)
- Think-as-You-See: Streaming Chain-of-Thought Reasoning for Large Vision-Language Models (CVPR 2026 · similarity 0.89)
- Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition (ICML 2024 · 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.
