Research map / Video / Video-language understanding
VidChain: Chain-of-Tasks with Metric-based Direct Preference Optimization for Dense Video Captioning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes VidChain, combining Chain-of-Tasks multi-turn decomposition of dense video captioning with metric-based DPO that uses SODA/METEOR/IoU as preference signals, improving VTimeLLM and VideoLLaMA2 on ActivityNet/YouCook2 DVC and ActivityNet TVG.
Problem
VideoLLMs perform poorly on fine-grained temporal understanding such as dense video captioning (DVC), which requires segmentation, captioning and temporal grounding; single-step generation and next-token training do not match DVC evaluation metrics.
Approach
Decompose DVC into sequential sub-tasks (segment count, timestamps, captions) asked in multi-turn QA in two orderings, then align the model with each sub-task's evaluation metric via DPO on preference pairs built from metric values (continuous preferences), with a margin threshold gamma that filters pairs with small metric gaps.
Key results
- VideoLLaMA2+VidChain (P_c→t) reaches SODA_c 8.8 on ActivityNet vs 7.2 baseline and CIDEr 43.9 vs 32.9.
- VTimeLLM+VidChain (7B) outperforms VTimeLLM 13B on DVC and TVG; on ActivityNet TVG VideoLLaMA2+VidChain gets R@0.3 63.3 vs 49.4 baseline.
- Ablation: M-DPO with gap threshold gamma (L_M-DPO) gives SODA_c 8.8 vs 8.3 for standard DPO on VideoLLaMA2+CoTasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CoTasks and M-DPO using the caption-then-timestamp path | VideoLLaMA2 · ActivityNet Captions val2 | SODA_c: 7.2 -> 8.8 | Helped |
| Add CoTasks and M-DPO using the caption-then-timestamp path | VideoLLaMA2 · ActivityNet Captions val2 | METEOR: 7.7 -> 8.8 | Helped |
| Add CoTasks and M-DPO using the caption-then-timestamp path | VideoLLaMA2 · ActivityNet Captions val2 | CIDEr: 32.9 -> 43.9 | Helped |
| Add CoTasks and M-DPO using the caption-then-timestamp path | VideoLLaMA2 · YouCook2 validation | SODA_c: 3.3 -> 4.8 | Helped |
| Add CoTasks and M-DPO using the caption-then-timestamp path | VideoLLaMA2 · YouCook2 validation | METEOR: 3.5 -> 5.6 | Helped |
| Add CoTasks and M-DPO using the caption-then-timestamp path | VideoLLaMA2 · YouCook2 validation | CIDEr: 12.3 -> 23.8 | Helped |
| Add CoTasks and M-DPO | VideoLLaMA2 · ActivityNet Captions val2 TVG | R@0.3: 49.4 -> 63.3 | Helped |
| Add CoTasks and M-DPO | VideoLLaMA2 · ActivityNet Captions val2 TVG | R@0.5: 26.8 -> 44.8 | Helped |
| Add CoTasks and M-DPO | VideoLLaMA2 · ActivityNet Captions val2 TVG | R@0.7: 15.0 -> 25.2 | Helped |
| Add CoTasks and M-DPO | VideoLLaMA2 · ActivityNet Captions val2 TVG | mIoU: 33.9 -> 44.1 | Helped |
| Add CoTasks without M-DPO (ablation) | VideoLLaMA2 · ActivityNet Captions val2 | SODA_c: 7.2 -> 7.7 | Helped |
| Add M-DPO to CoTasks (ablation) | VideoLLaMA2 · ActivityNet Captions val2 | SODA_c: 7.7 -> 8.8 | Helped |
Surprising findings
- Finetuning with next-token loss on preferred samples only yields little or inconsistent gain (SODA_c drops 0.1 for VideoLLaMA2+CoTasks-P_t→c), while M-DPO gives gains.
- Using SODA as the sole metric for sub-task preference pairs is noted as equivalent to METEOR for captioning sub-tasks, simplifying construction.
Limitations
- Gains depend on the choice of inference path; P_c→t vs P_t→c trade-offs vary across models.
- Inference cost increases due to multi-turn decomposition.
- Some videos are no longer available online, so a few training samples are missing.
Experimental setup
Datasets: ActivityNet Captions YouCook2
Benchmarks: ActivityNet Captions YouCook2
Metrics: SODA_c METEOR CIDEr R@0.3 R@0.5 R@0.7 mIoU
Baselines: VideoChat VideoLLaMA VideoChatGPT TimeChat VTimeLLM VideoLLaMA2 PDVC
Builds on: DPO Step-DPO VTimeLLM VideoLLaMA2
Closest papers by meaning
- LongVPO: From Anchored Cues to Self-Reasoning for Long-Form Video Preference Optimization (NeurIPS 2025 · similarity 0.89)
- Self-alignment of Large Video Language Models with Refined Regularized Preference Optimization (NeurIPS 2025 · similarity 0.88)
- Dense Video Object Captioning from Disjoint Supervision (ICLR 2025 · similarity 0.88)
- HiCM²: Hierarchical Compact Memory Modeling for Dense Video Captioning (AAAI 2025 · similarity 0.88)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.88)
- Distilling Vision-Language Models on Millions of Videos (CVPR 2024 · 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.