--- base_model: - Qwen/Qwen2.5-VL-7B-Instruct datasets: - OX-PIXL/STVQA-7K license: apache-2.0 library_name: transformers pipeline_tag: image-text-to-text --- # SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards

Paper Project Page GitHub Repo Hugging Face Models

## Model Description 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. 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.

SpatialThinker Overview

## Model Details * **Developed by:** Hunar Batra, Haoqin Tu, Hardy Chen, Yuanze Lin, Cihang Xie, Ronald Clark * **Model type:** 3D-aware Multimodal Large Language Model (MLLM) * **Language(s) (NLP):** English * **License:** Apache-2.0 * **Finetuned from model:** [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) ### Model Sources * **Repository:** [https://github.com/hunarbatra/SpatialThinker](https://github.com/hunarbatra/SpatialThinker) * **Paper:** [https://huggingface.co/papers/2511.07403](https://huggingface.co/papers/2511.07403) * **Project Page:** [https://hunarbatra.com/SpatialThinker/](https://hunarbatra.com/SpatialThinker/) ## How to Get Started with the Model This model can be loaded and used directly with the Hugging Face `transformers` library. First, ensure you have the necessary dependencies installed: ```bash pip install transformers>=4.49.0 pip install flash-attn>=2.4.3 vllm>=0.7.3 # (vllm 0.8.0 recommended) ``` Then, you can use the following Python code snippet for inference: ```python import torch from transformers import AutoProcessor, AutoModelForCausalLM from PIL import Image import requests from io import BytesIO # Load model and processor model_id = "OX-PIXL/SpatialThinker-7B" # This is the model repository ID processor = AutoProcessor.from_pretrained(model_id) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.bfloat16, # Use bfloat16 for better performance on compatible GPUs device_map="auto", trust_remote_code=True # Required for custom Qwen2.5-VL architecture ).eval() # Set model to evaluation mode # Example image (replace with your own image path or URL) image_url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg" response = requests.get(image_url) image = Image.open(BytesIO(response.content)).convert("RGB") # Define a spatial reasoning question question = "What are the spatial relationships between the car, the road, and the trees?" # Construct chat messages messages = [ {"role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": question} ]} ] # Apply chat template and process inputs prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device) # Generate response with torch.no_grad(): output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False) # Use suitable generation parameters response_text = processor.decode(output_ids[0], skip_special_tokens=True) print(f"Question: {question} ") print(f"Answer: {response_text}") ``` ## Updates * **[2025/11/11]** 🔥 Code base released. * **[2025/11/08]** 🔥 Model Checkpoints and Dataset released. ## 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 Details ### Training Procedure SpatialThinker models are trained with STVQA-7K, Dense Spatial Rewards + GRPO. Baseline models (Vanilla GRPO) are also trained with STVQA-7K. #### Train **SpatialThinker Models** with STVQA-7K, Dense Spatial Rewards + GRPO ```bash bash scripts/spatialthinker_3b_grpo.sh 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 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 \ --template \ # e.g. `reasoning`, `no_reasoning`, `spatial_thinker` --model_path \ --cuda \ --batch_size \ [--provider ] \ [--processor_name ] \ [--custom_filename ] ``` ### 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