tsinghua-ee/D-ORCA-8B-0210

D-ORCA: Dialogue-Centric Optimization for Robust Audio-Visual Captioning

Resolving the Triplet of "When, Who, and What is Said" in Dialogue-Centric Videos

---

🔥 News

• [2026/02/10] 🚀 D-ORCA inference code and the model checkpoint are released! We achieve SOTA results on dialogue-centric video understanding among open-sourced models.
• [2026/02/08] 📄 Our paper is available on arXiv.

📖 Introduction

We introduce D-ORCA, a dialogue-centric omni-modal large language model optimized for robust audio-visual captioning. Unlike existing models that struggle with speaker attribution and temporal alignment, D-ORCA is designed to accurately resolve when, who, and what is said in the video.

Most open-source multimodal LLMs fail to produce accurate, dialogue-aware captions (see Figure 1). D-ORCA addresses this by:

1. Constructing DVD-Train: A large-scale (40k videos) bilingual dataset tailored for dialogue scenarios.
2. Novel RL Optimization: Adopting GRPO with three specialized rewards:
   • 🎯 Speaker Attribution Accuracy
   • 📝 Global Speech Content Accuracy
   • ⏱️ Sentence-level Temporal Boundary Alignment

Despite having only 8B parameters, D-ORCA outperforms existing open-source models and remains competitive with significantly larger models on general benchmarks.

Figure 1: Comparison of D-ORCA with other models. D-ORCA accurately identifies speakers, recognizes speech, and aligns timestamps.

🏆 Performance

D-ORCA achieves state-of-the-art performance on our curated DVD-Bench.

Model	(En) Acc% ↑	(En) WER% ↓	(En) IoU% ↑	(Zh) Acc% ↑	(Zh) WER% ↓	(Zh) IoU% ↑
ARC-Qwen-Video-Narrator (7B)	66.4	65.0	23.0	63.2	53.6	10.1
Qwen2.5-Omni (7B)	62.7	83.6	-	55.7	69.4	-
video-SALMONN 2+ (7B)	66.6	94.0	-	59.9	-	-
AVoCaDO (7B)	72.9	17.9	-	69.3	-	-
Qwen3-Omni-Instruct (30B)	67.8	91.3	-	63.5	60.6	-
D-ORCA (8B)	81.1	16.6	57.1	78.0	17.5	37.8

D-ORCA also achieves competitive results on general audio-visual benchmarks.

Model	Video-MME	WorldSense	AVUT	Video-Holmes	DailyOmni	AV-SpeakerBench
OmniVinci (7B)	68.6	48.2	-	-	66.5	-
ARC-Qwen-Video-Narrator (7B)	62.4	45.1	-	43.2	63.2	40.2
Qwen2.5-Omni (7B)	64.3	47.8	66.3	43.7	62.7	42.3
video-SALMONN 2+ (7B)	73.4	50.9	69.5	46.9	71.8	51.6
AVoCaDO (7B)	65.9	49.9	70.0	47.2	69.8	51.6
Qwen3-Omni-Instruct (30B-A3B)	70.5	54.0	72.0	54.1	69.8	54.1
D-ORCA (8B)	72.9	53.7	76.1	48.5	78.5	55.0

🛠️ Quick Start

Model Zoo

Model	Base LLM	Params	HuggingFace
D-ORCA-8B-0210	Qwen3-VL-8B	8B	Download

Installation

For key library versions, please refer to new_req.txt. Install via:

pip install -r new_req.txt

Inference

For evaluation on a whole dataset:

1. Prepare the dataset following scripts/example_data.json.
2. Modify the parameters in scripts/direct_test.sh.
3. Run bash scripts/direct_test.sh.

For setting up a CLI demo:

1. Modify the parameters in scripts/direct_demo.sh.
2. Run bash scripts/direct_demo.sh.
3. Modify scripts/demo_config.yaml to control input.

📅 Roadmap

• Release D-ORCA 8B model checkpoints.
• Release Inference Code.
• Release DVD-Bench evaluation data.
• Release Training Code (SFT, pre-DPO, GRPO).
• Release DVD-Train dataset annotations.

🖊️ Citation

If you find D-ORCA useful for your research, please cite our paper:

@article{tang2026dorca,
  title={{D-ORCA: Dialogue-Centric Optimization for Robust Audio-Visual Captioning}}, 
  author={Changli Tang and Tianyi Wang and Fengyun Rao and Jing LYU and Chao Zhang},
  journal={arXiv preprint arXiv:2602.07960},
  year={2026}
}

📄 License

This project is licensed under the Apache 2.0 License.