Research map / Video / Video-language understanding
Towards Effective Long Video Understanding: Dynamic MAS Construction via Meta-Agent
Jing Huang, Lidong Zhang, Yadong Li, Xingzhong Xu, Siye Chen, Jie Liu, Ming Kong, Qiang Zhu
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 |
|---|---|---|---|
| add SFT on synthesized MAS code (ablation) | DyMAC MAS-Visual · MLVU/Video-MME/LVB average | accuracy: 51.2 -> 65.5 | Improved |
| add RLVR after SFT (ablation) | DyMAC MAS-Visual · MLVU/Video-MME/LVB average | accuracy: 65.5 -> 68.2 | Improved |
| give Meta-Agent the video as well as the query (ablation) | DyMAC after RL · MLVU/Video-MME/LVB average | accuracy: 64.4 -> 68.2 | Improved |
| add SFT to text-only Meta-Agent (ablation) | DyMAC MAS-Text · MLVU/Video-MME/LVB average | accuracy: 50.9 -> 63.1 | Improved |
| untrained dynamic MAS vs vanilla baseline (ablation) | DyMAC MAS-Visual · MLVU/Video-MME/LVB average | accuracy: 63.3 -> 51.2 | Hurt |
| add accuracy reward in RLVR (ablation) | DyMAC · MLVU | accuracy: 70.9 -> 72.7 | Improved |
| add efficiency reward to accuracy reward (ablation) | DyMAC · Inference time | seconds: 13.2 -> 4.6 | Improved |
| add efficiency reward to accuracy reward (ablation) | DyMAC · MLVU | accuracy: 72.7 -> 73.2 | No clear effect |
| dynamic 3b/7b driving model allocation instead of 3b only (ablation) | DyMAC · MLVU | accuracy: 69.7 -> 73.2 | Improved |
| dynamic 3b/7b allocation instead of 7b only (ablation) | DyMAC · Inference time | seconds: 9.8 -> 4.6 | Improved |
| add CoT to reasoning patterns (ablation) | DyMAC · MLVU/Video-MME/LVB average | accuracy: 64.5 -> 65.7 | Improved |
| add LLM-Debate to reasoning patterns (ablation) | DyMAC · MLVU/Video-MME/LVB average | accuracy: 65.7 -> 66.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- Video-o3 — Video-o3: Native Interleaved Clue Seeking for Long Video Multi-Hop Reasoning (ICML 2026) · better on 4 of 4 shared results
- VideoLucy — VideoLucy: Deep Memory Backtracking for Long Video Understanding (NeurIPS 2025) · better on 0 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Think, Then Verify: A Hypothesis-Verification Multi-Agent Framework for Long Video Understanding (CVPR 2026 · similarity 0.91)
- Deep Video Discovery: Agentic Search with Tool Use for Long-form Video Understanding (NeurIPS 2025 · similarity 0.89)
- Enhancing Video-LLM Reasoning via Agent-of-Thoughts Distillation (CVPR 2025 · similarity 0.89)
- ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding (NeurIPS 2025 · similarity 0.88)
- Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA (NeurIPS 2026 · similarity 0.88)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · 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.
