---
base_model:
- Qwen/Qwen2.5-VL-3B-Instruct
datasets:
- OX-PIXL/STVQA-7K
pipeline_tag: image-text-to-text
library_name: transformers
---

# SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards

The `SpatialThinker` model, presented in the paper [SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards](https://huggingface.co/papers/2511.07403), is a 3D-aware Multimodal Large Language Model (MLLM) trained with Reinforcement Learning (RL) to integrate structured spatial grounding with multi-step reasoning.

**Paper (ArXiv)**: [https://arxiv.org/abs/2511.07403](https://arxiv.org/abs/2511.07403)
**Project Page**: [https://hunarbatra.com/SpatialThinker/](https://hunarbatra.com/SpatialThinker/)
**Code**: [https://github.com/hunarbatra/SpatialThinker](https://github.com/hunarbatra/SpatialThinker)

## Abstract
Multimodal large language models (MLLMs) have achieved remarkable progress in vision-language tasks, but they continue to struggle with spatial understanding. Existing spatial MLLMs often rely on explicit 3D inputs or architecture-specific modifications, and remain constrained by large-scale datasets or sparse supervision. To address these limitations, we introduce SpatialThinker, a 3D-aware MLLM trained with RL to integrate structured spatial grounding with multi-step reasoning. The model simulates human-like spatial perception by constructing a scene graph of task-relevant objects and spatial relations, and reasoning towards an answer via dense spatial rewards. SpatialThinker consists of two key contributions: (1) a data synthesis pipeline that generates STVQA-7K, a high-quality spatial VQA dataset, and (2) online RL with a multi-objective dense spatial reward enforcing spatial grounding. SpatialThinker-7B outperforms supervised fine-tuning and the sparse RL baseline on spatial understanding and real-world VQA benchmarks, nearly doubling the base-model gain compared to sparse RL, and surpassing GPT-4o. These results showcase the effectiveness of combining spatial supervision with reward-aligned reasoning in enabling robust 3D spatial understanding with limited data and advancing MLLMs towards human-level visual reasoning.

<p align="center">
  <img src="https://github.com/hunarbatra/SpatialThinker/raw/main/assets/spatialthinker.jpg" width="60%" alt="SpatialThinker Overview">
</p>

---

### 🧩 Requirements

- Python 3.9+
- `transformers >= 4.49.0`
- `flash-attn >= 2.4.3`
- `vllm >= 0.7.3` (0.8.0 recommended)

---

### ⚙️ Installation

```bash
pip install -e .
```

---

### 🚀 Training

#### Train **SpatialThinker Models** with STVQA-7K, Dense Spatial Rewards + GRPO

```bash
bash scripts/spatialthinker_3b_grpo.sh
```
```bash
bash scripts/spatialthinker_7b_grpo.sh
```

#### Train **Baseline Models** (Vanilla GRPO) with STVQA-7K

```bash
bash scripts/qwen_2_5_3b_stvqa_vanilla_grpo.sh
```
```bash
bash scripts/qwen_2_5_7b_stvqa_vanilla_grpo.sh
```

---

### 🧠 Merge Checkpoints to Hugging Face Format
```bash
python3 scripts/model_merger.py --local_dir path_to_your_last_actor_checkpoint
```
---

### 🧪 Evaluation

To evaluate **SpatialThinker** or baseline models across spatial reasoning benchmarks, use the provided `evaluation/eval.py` script.

#### Basic Command Structure
```bash
python3 evaluation/eval.py \
    --dataset <dataset_name> \
    --template <prompt_template> \ # e.g. `reasoning`, `no_reasoning`, `spatial_thinker`  
    --model_path <model_or_checkpoint> \
    --cuda <gpu_id> \
    --batch_size <num_samples_per_step> \
    [--provider <inference_backend>] \ 
    [--processor_name <tokenizer_or_processor>] \
    [--custom_filename <output_name>]
```

#### ⚙️ Example: Evaluate Across Multiple Benchmarks

```bash
python3 evaluation/eval.py \
    --dataset blink-spatial \
    --template spatial_thinker \
    --model_path OX-PIXL/SpatialThinker-3B \
    --cuda 0 \
    --batch_size 4
```
```bash
python3 evaluation/eval.py \
    --dataset spatialbench \
    --template spatial_thinker \
    --model_path OX-PIXL/SpatialThinker-3B \
    --cuda 0 \
    --batch_size 2
```

#### 📊 Example: Evaluate Using an API Provider (OpenAI / Anthropic)

```bash
python3 evaluation/eval.py \
    --dataset stvqa \
    --template reasoning \
    --model_path gpt-4o-2024-05-13 \
    --provider openai \
    --batch_size 1
```
```bash
python3 evaluation/eval.py \
    --dataset stvqa \
    --template reasoning \
    --model_path claude-3-5-sonnet \
    --provider anthropic \
    --batch_size 1
```

#### Supported Evaluation Datasets
`cv-bench`, `cv-bench-2D`, `cv-bench-3D`, `blink-spatial`, `blink-depth`, `blink-object`,  
`blink-counting`, `blink-multi-view`, `blink-jigsaw`, `realworld_qa`, `spatialbench`, `mmvp`, `3dsrbench`,
`lego`, `spatialreasoner`, `robospatial`, `robospatial_rgb`, `stvqa`, `hallusionbench`.

---
### 📘 Citation

If you find this repository useful in your project, please consider giving a ⭐ and citing:

```bibtex
@misc{batra2025spatialthinkerreinforcing3dreasoning,  
 title={SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards},  
 author={Hunar Batra and Haoqin Tu and Hardy Chen and Yuanze Lin and Cihang Xie and Ronald Clark},  
 year={2025},  
 eprint={2511.07403},  
 archivePrefix={arXiv},  
 primaryClass={cs.CV},  
 url={https://arxiv.org/abs/2511.07403},  
}
```
---

### 🌟 Acknowledgements
This project builds upon the following open-source frameworks and works:
- [**EasyR1**](https://github.com/hiyouga/EasyR1) — An efficient, scalable, multi-modality RL training framework based on veRL  
- [**LLaMA-Factory**](https://github.com/hunarbatra/LLaMA-Factory) — Unified efficient fine-tuning of 100+ LLMs & VLMs  
- [**Qwen2.5-VL**](https://arxiv.org/abs/2502.13923) — Multimodal LLM series from the Qwen family
  
---