Research map / Video / Video-language understanding
META: Meta Evolution of Tool Trajectory Adaptation for Long-Video Understanding
Jing Huang, Luyuan Chen, Zhijie Xu, 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 Orchestrator-Worker framework (ablation) | META · MLVU / Video-MME (Qwen3-VL-8B) | accuracy: 78.1 -> 80.8 (MLVU), 71.4 -> 73.5 (Video-MME) | Improved |
| add micro tool evolution (ablation) | META · Video-MME | accuracy: 73.5 -> 75.7 | Improved |
| add macro tool evolution (ablation) | META · Video-MME | accuracy: 73.5 -> 76.6 | Improved |
| add both micro and macro tool evolution (ablation) | META · MLVU | accuracy: 80.8 -> 83.6 | Improved |
| increase evolution iterations from 1 to 3 (ablation) | META · MLVU | accuracy: 79.7 -> 83.6 | Improved |
| increase evolution iterations from 3 to 5 (ablation) | META · MLVU | accuracy: 83.6 -> 83.8 | No clear effect |
| use weaker LLM (Qwen3-8B) for Reflector (ablation) | META · MLVU | accuracy: 83.6 -> 81.6 | Hurt |
| use weaker LLM for Orchestrator-Worker (ablation) | META · MLVU | accuracy: 83.6 -> 81.2 | Hurt |
| evolve without ground-truth labels (ablation) | META · Video-MME (Qwen3-VL-8B) | accuracy: 77.7 -> 76.5 | No clear effect |
| increase evolution samples from 100 to 1000 (ablation) | META · MLVU (Qwen2.5-VL-7B, offline) | accuracy: 70.6 -> 77.3 | Improved |
| equip ReAct agent with evolved toolset (ablation) | ReAct · Video-MME | accuracy: 69.8 -> 72.5 | Improved |
| equip Plan&Execute agent with evolved toolset (ablation) | Plan&Execute · MLVU | accuracy: 79.0 -> 81.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VideoLucy — VideoLucy: Deep Memory Backtracking for Long Video Understanding (NeurIPS 2025) · better on 3 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
- LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding (CVPR 2026 · similarity 0.90)
- VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning (ICLR 2026 · similarity 0.90)
- ROVER: Recursive Reasoning Over Videos with Vision-Language Models for Embodied Tasks (NeurIPS 2025 · similarity 0.88)
- Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.88)
- AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning (ICLR 2026 · similarity 0.88)
- Mimic Human Cognition, Master Multi-Image Reasoning: A Meta-Action Framework for Enhanced Visual Understanding (CVPR 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.
