Research map / Multi-agent systems and games / LLM agents
VDC-Agent: When Video Detailed Captioners Evolve Themselves via Agentic Self-Reflection
Qiang Wang, XINYUAN GAO, SongLin Dong, Jizhou Han, Jiangyang Li, Yuhang He, Zhiheng Ma, Yihong Gong
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 principle text to prompt (ablation) | Qwen2.5-VL-7B · VDC | average accuracy: 43.95 -> 46.47 | Improved |
| add test-time self-reflection on top of principles (ablation) | Qwen2.5-VL-7B · VDC | average accuracy: 46.47 -> 47.10 | Improved |
| add test-time self-reflection on top of principles (ablation) | Qwen2.5-VL-7B · VDC | inference time: 22.3s -> 164.9s | Hurt |
| internalize reflection by curriculum DPO training instead of test-time reflection (ablation) | Qwen2.5-VL-7B · VDC | average accuracy: 47.10 -> 49.08 | Improved |
| use curriculum (easy-to-hard) DPO instead of random order (ablation) | VDC-Agent · VDC | average accuracy: 48.03 -> 49.08 | Improved |
| use hard-to-easy order instead of random order (ablation) | VDC-Agent · VDC | average accuracy: 48.03 -> 47.65 | Hurt |
| use DPO curriculum instead of SFT (ablation) | VDC-Agent · VDC | average accuracy: 47.54 -> 49.08 | Improved |
| raise max iterations T from 4 to 6 (ablation) | VDC-Agent · VDC | average accuracy: 49.08 -> 49.55 | No clear effect |
| raise max iterations T from 2 to 4 (ablation) | VDC-Agent · DREAM-1K | F1: 34.0 -> 37.4 | Improved |
| swap principle set P1 for P2-P4 written by others (ablation) | VDC-Agent · VDC | average accuracy: 49.08 vs 48.84, 49.02, 48.95 | No clear effect |
| apply VDC-Agent to LLaVA-OneVision-7B | LLaVA-OneVision-7B · VDC | average accuracy: 37.45 -> 43.18 | Improved |
| apply VDC-Agent to InternVL2.5-8B | InternVL2.5-8B · DREAM-1K | F1: 27.6 -> 33.9 | Improved |
Closest papers by meaning
- MM-SeR: Multimodal Self-Refinement for Lightweight Image Captioning (CVPR 2026 · similarity 0.87)
- AuroraCap: Efficient, Performant Video Detailed Captioning and a New Benchmark (ICLR 2025 · similarity 0.87)
- ESCA: Contextualizing Embodied Agents via Scene-Graph Generation (NeurIPS 2025 · similarity 0.86)
- MetaCaptioner: Towards Generalist Visual Captioning with Open-source Suites (ICLR 2026 · similarity 0.86)
- Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.86)
- ViDiC: Video Difference Captioning (NeurIPS 2026 · similarity 0.86)
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.
