Research map / Video / Video-language understanding
VidBridge-R1: Bridging QA and Captioning for RL-based Video Understanding Models with Intermediate Proxy Tasks
Xinlong Chen, Yuanxing Zhang, Yushuo Guan, Weihong Lin, Zekun Moore Wang, Bohan Zeng, Yang Shi, Sihan Yang, Qiang Liu, Pengfei Wan, Liang 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 |
|---|---|---|---|
| naively mix VideoQA and Caption data and rewards in RL (ablation) | Qwen2.5-VL-7B GRPO · Video-MME | accuracy: 63.2 -> 54.8 | Hurt |
| naively mix VideoQA and Caption in RL, effect on captioning (ablation) | Qwen2.5-VL-7B GRPO · DREAM-1K | AutoDQ F1: 34.8 -> 30.6 | Hurt |
| add DarkEventInfer proxy task to VideoQA (ablation) | RL training mix · DarkEventInfer-Test / MixVidQA-Test | judge score / accuracy: 60.0 -> 113.0 / 24.0 -> 41.0 | Improved |
| add MixVidQA proxy task on top of VideoQA + DarkEventInfer (ablation) | RL training mix · MixVidQA-Test | accuracy: 41.0 -> 54.0 | Improved |
| add Caption task on top of VideoQA + both proxy tasks (ablation) | RL training mix · DREAM-1K | AutoDQ F1: 32.2 -> 35.2 | Improved |
| add Caption task on top of VideoQA + both proxy tasks, effect on reasoning (ablation) | RL training mix · DarkEventInfer-Test | judge score: 121.0 -> 117.0 | Hurt |
| scale up VideoQA and Caption data volume in naive mix (ablation) | RL training mix · Video-MME | accuracy: 54.8 -> 61.5 | Mixed |
| add KL divergence regularization to GRPO (coefficient 0.10) (ablation) | VidBridge-R1 · Video-MME | accuracy: 64.3 -> 62.6 | Hurt |
| use implicit format reward instead of explicit or none (ablation) | VidBridge-R1 · MVBench | accuracy: 58.2 -> 61.9 | Improved |
| set precision weight alpha=1.0 without keywords reward in caption reward (ablation) | caption reward · VidCapBench | accuracy: 13.0 -> 10.3 | Hurt |
| add keywords reward (beta=0.2) to caption reward (ablation) | caption reward · DREAM-1K | AutoDQ F1: 32.9 -> 34.8 | Improved |
| use SFT cold start with reasoning chains before RL (ablation) | Qwen2.5-VL-7B · Video-MME | accuracy: 57.0 vs 64.3 for direct RL | Hurt |
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)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.90)
- MoReVQA: Exploring Modular Reasoning Models for Video Question Answering (CVPR 2024 · similarity 0.90)
- Video-R1: Reinforcing Video Reasoning in MLLMs (NeurIPS 2025 · similarity 0.90)
- ReWatch-R1: Boosting Complex Video Reasoning in Large Vision-Language Models through Agentic Data Synthesis (ICLR 2026 · similarity 0.89)
- CG-Bench: Clue-grounded Question Answering Benchmark for Long Video Understanding (ICLR 2025 · 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.
