Research map / Video / Video-language understanding
DiaDem: Advancing Dialogue Descriptions in Audiovisual Video Captioning for Multimodal Large Language Models
Xinlong Chen, Weihong Lin, Jingyun Hua, Linli Yao, Yue Ding, Bozhou Li, Bohan Zeng, Yang Shi, Qiang Liu, Yuanxing Zhang, 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 |
|---|---|---|---|
| add SFT on 85K synthesized dialogue-aware captions (ablation) | AVoCaDO · DiaDemBench (overall) | REF / ASR: 38.7 / 51.7 -> 59.3 / 74.0 | Improved |
| add SFT on 85K synthesized captions (ablation) | AVoCaDO · video-SALMONN-2 testset | Total error: 37.3 -> 42.0 | Hurt |
| add GRPO stage 1 after SFT (ablation) | DiaDem · DiaDemBench (overall) | REF / ASR: 59.3 / 74.0 -> 64.7 / 77.6 | Improved |
| add GRPO stage 2 on duplicated high-difficulty subset (ablation) | DiaDem · DiaDemBench (overall) | REF / ASR: 64.7 / 77.6 -> 65.9 / 79.3 | Improved |
| add GRPO after SFT to fix SALMONN-2 degradation (ablation) | DiaDem · video-SALMONN-2 testset | Total error: 42.0 -> 36.5 | Improved |
| use staged training instead of single merged phase (ablation) | GRPO · DiaDemBench (overall) | REF / ASR: 64.5 / 78.2 -> 65.9 / 79.3 | Improved |
| add easy-sample filtering (ablation) | GRPO · DiaDemBench (overall) | REF / ASR: 63.2 / 78.2 -> 65.9 / 79.3 | Improved |
| add the dialogue reward (ablation) | GRPO · DiaDemBench (overall) | REF / ASR: 60.1 / 74.3 -> 65.9 / 79.3 | Improved |
| add dialogue reward (ablation) | GRPO · UGC-VideoCap | Avg. score: 75.2 -> 75.6 | No clear effect |
| use GRPO data as extra SFT data instead of RL (ablation) | SFT model · DiaDemBench (overall) | REF / ASR: 59.3 / 74.0 -> 59.2 / 73.6 | No clear effect |
| replace AVoCaDO dialogue reward with enhanced dialogue reward (ablation) | GRPO Stage-2 · DiaDemBench (overall) | REF / ASR: 63.7 / 77.2 -> 65.9 / 79.3 | Improved |
| post-train AVoCaDO into DiaDem | AVoCaDO · UGC-VideoCap | Avg. score: 73.2 -> 75.6 | 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.
- AVoCaDO — AVoCaDO: An Audiovisual Video Captioner Driven by Temporal Orchestration (ICLR 2026) · better on 4 of 4 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
- OmniCapBench: A Deep-Structured Evaluation Framework for Fine-Grained Audio-Visual Captioning (NeurIPS 2026 · similarity 0.88)
- Distilling Vision-Language Models on Millions of Videos (CVPR 2024 · similarity 0.87)
- CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning (ICLR 2026 · similarity 0.87)
- Vript: A Video Is Worth Thousands of Words (NeurIPS 2024 · similarity 0.86)
- F2RVLM: Boosting Fine-grained Fragment Retrieval for Multi-Modal Long-form Dialogue with Vision Language Model (AAAI 2026 · similarity 0.86)
- V^2Dial: Unification of Video and Visual Dialog via Multimodal Experts (CVPR 2025 · 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.
