Research map / Video / Video-language understanding
Clue Matters: Empower Video Reasoning with Brain-Inspired Latent Clue Learning
Kaixin Zhang, Xiaohe Li, Jiahao Li, Haohua Wu, Xinyu Zhao, Zide Fan, Lei 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 |
|---|---|---|---|
| train Stage 1 decoupled supervision only (without Stage 2) (ablation) | VideoLLaMA3 · STAR (total) | accuracy: 72.1 -> 70.8 | Hurt |
| add Stage 2 inference supervision after Stage 1 (ablation) | ClueNet · STAR (total) | accuracy: 70.8 -> 75.1 | Improved |
| add Stage 2 inference supervision after Stage 1 (ablation) | ClueNet · NExT-QA MC (total) | accuracy: 80.6 -> 83.3 | Improved |
| add Keyframe Selection (ablation) | ClueNet · STAR (total) | accuracy: 75.1 -> 75.4 | No clear effect |
| add Adaptive Clue Filter (ablation) | ClueNet · STAR (total) | accuracy: 75.4 -> 77.6 | Improved |
| add Adaptive Clue Filter (ablation) | ClueNet · NExT-QA MC (total) | accuracy: 83.4 -> 84.3 | Improved |
| add Visual Compression (ablation) | ClueNet · Inference GFLOPs | GFLOPs: 51,355.89 -> 40,622.84 | Improved |
| add Visual Compression (ablation) | ClueNet · STAR (total) | accuracy: 77.6 -> 77.3 | No clear effect |
| skip Stage 1 training (ablation) | ClueNet · STAR (total) | accuracy: -> 33.1 | Hurt |
| supply ground-truth clues at inference (oracle) | Qwen2.5-VL · STAR oracle clues | accuracy: +22.9% | Improved |
| replace VideoLLaMA3 with ClueNet | VideoLLaMA3 · STAR (Feasibility) | accuracy: 68.0 vs 66.7 | Hurt |
| apply ClueNet clue-based reasoning on VideoLLaMA3 | VideoLLaMA3 · NExT-OE (descriptive) | WUPS@0: 54.6 -> 64.9 vs Qwen2.5-VL | Improved |
Closest papers by meaning
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.90)
- CG-Bench: Clue-grounded Question Answering Benchmark for Long Video Understanding (ICLR 2025 · similarity 0.90)
- Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation (CVPR 2025 · similarity 0.89)
- Divide and Conquer: Exploring Language-centric Tree Reasoning for Video Question-Answering (ICML 2025 · similarity 0.89)
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · similarity 0.89)
- MoReVQA: Exploring Modular Reasoning Models for Video Question Answering (CVPR 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.
