Research map / Video / Video-language understanding
Fostering Video Reasoning via Next-Event Prediction
Haonan Wang, Hongfu Liu, Xiangyan Liu, Chao Du, Kenji Kawaguchi, Ye Wang, Tianyu Pang
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 |
|---|---|---|---|
| train on next-event prediction instead of captioning/MCQA/OEQA | Qwen2.5-VL-7B-Instruct · Temporal benchmarks average | accuracy: 49.7 -> 53.5 | Improved |
| train on NEP | Qwen2.5-VL-7B-Instruct · General benchmarks average | accuracy: 60.3 -> 60.9 | No clear effect |
| use next-event (deductive) instead of video QA or previous-event prediction (ablation) | Qwen2.5-VL-7B-Instruct · FutureBench | accuracy: 58.8 (QA) / 55.1 (PEP) -> 61.3 | Improved |
| use previous-event prediction instead of NEP (ablation) | Qwen2.5-VL-7B-Instruct · SeedBench-R1 | accuracy: 39.5 -> 31.2 | Hurt |
| use Mix tuning instead of SFT (ablation) | Qwen2.5-VL-7B-Instruct · Temporal benchmarks average | accuracy: 52.6 -> 53.3 | Improved |
| use Distill instead of SFT (ablation) | Qwen2.5-VL-7B-Instruct · General benchmarks average | accuracy: 59.7 -> 61.2 | Improved |
| use critique fine-tuning instead of SFT (ablation) | Qwen2.5-VL-7B-Instruct · FutureBench | accuracy: 61.3 -> 55.8 | Hurt |
| increase NEP training data beyond 5K samples (ablation) | SFT/CFT/Distill/Mix · General and temporal benchmarks | accuracy | No clear effect |
| train with GRPO instead of SFT on NEP-style MCQA (ablation) | Qwen2.5-VL-7B-Instruct · FutureBench 1-Hop | accuracy: 67.6 -> 83.8 | Improved |
| train with GRPO instead of SFT on NEP-style MCQA (ablation) | Qwen2.5-VL-7B-Instruct · FutureBench 3-Hop (OOD) | accuracy: 57.7 -> 62.7 | Improved |
| train with GRPO instead of SFT (ablation) | Qwen2.5-VL-7B-Instruct · General benchmarks average | accuracy: 59.7 -> 58.2 | Hurt |
| train with captioning on full video | Qwen2.5-VL-7B-Instruct · TempCompass | accuracy: 73.8 -> 72.2 | Hurt |
Closest papers by meaning
- EventLens: Event-Structure Reinforcement Learning for Video Understanding (NeurIPS 2026 · similarity 0.88)
- Momentor: Advancing Video Large Language Model with Fine-Grained Temporal Reasoning (ICML 2024 · similarity 0.87)
- VTimeLLM: Empower LLM to Grasp Video Moments (CVPR 2024 · similarity 0.87)
- Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs (NeurIPS 2026 · similarity 0.87)
- VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning (NeurIPS 2025 · similarity 0.87)
- Temporal Reasoning Transfer from Text to Video (ICLR 2025 · similarity 0.87)
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.
