---
license: apache-2.0
base_model: Qwen/Qwen2.5-VL-3B-Instruct
base_model_relation: finetune
library_name: transformers
pipeline_tag: image-text-to-text
language:
- en
datasets:
- MLL-Lab/MindCube
tags:
- multi-view
- spatial-reasoning
- 3d-reasoning
- visual-question-answering
- cognitive-map
---

# DR-MV3D (SFT)

Official checkpoint for **[Dense Reward for Multi-View 3D Reasoning with Global Maps and Local Views](https://arxiv.org/abs/2606.23557)** (ECCV 2026).

[Paper](https://arxiv.org/abs/2606.23557) · [Project page](https://dr-mv3d.github.io/) · [Code](https://github.com/kaist-cvml/DR-MV3D) · [SFT + GRPO checkpoint](https://huggingface.co/jihochoi/DR-MV3D-R-GRPO)

DR-MV3D answers questions about a 3D scene shown from several viewpoints. This is the supervised fine-tuning stage: [Qwen2.5-VL-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct) trained on map-grounded reasoning traces. For general use prefer the [final SFT + GRPO checkpoint](https://huggingface.co/jihochoi/DR-MV3D-R-GRPO).

## Usage

The model expects the structured task prompt it was trained with, and produces a cognitive map, per-view ego-centric maps, a reasoning chain and the answer. See the [GitHub repository](https://github.com/kaist-cvml/DR-MV3D) for the prompt format, data preparation, inference and evaluation.

```bash
python scripts/run_inference.py \
  --model-path jihochoi/DR-MV3D-R-SFT \
  -i data/prompts/MindCube_tinybench_drmv3d.jsonl \
  -o results/tinybench_responses.jsonl \
  --image-root data --batch-size 8
```

## License

The weights are a fine-tune of Qwen2.5-VL-3B-Instruct and are released under Apache-2.0, the base model's license. The code is MIT-licensed.

## Citation

```bibtex
@inproceedings{drmv3d2026,
  title={Dense Reward for Multi-View 3D Reasoning with Global Maps and Local Views},
  author={Choi, Jiho and Lee, Seonho and Park, Seojeong and Shim, Hyunjung},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2026},
  eprint={2606.23557},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}
```