OX-PIXL/SpatialThinker-7B
SpatialThinker-7B by OX-PIXL is a 7 billion parameter 3D-aware multimodal large language model (MLLM) built on Qwen2.5-VL-7B-Instruct. It is specifically trained with reinforcement learning using GRPO and dense spatial rewards to integrate structured spatial grounding with multi-step reasoning. This model excels at simulating human-like spatial perception by constructing scene graphs and reasoning across nine spatial categories, outperforming GPT-4o on spatial understanding benchmarks.
Loading preview...
SpatialThinker-7B: 3D-Aware Multimodal Reasoning
SpatialThinker-7B, developed by Hunar Batra and researchers from the University of Oxford and UC Santa Cruz, is a 7 billion parameter multimodal large language model (MLLM) designed for advanced spatial understanding. Built upon the Qwen2.5-VL-7B-Instruct base, this model is uniquely trained using Group Relative Policy Optimization (GRPO) with dense spatial rewards to foster human-like spatial perception.
Key Capabilities
- Structured Spatial Reasoning: Constructs question-focused scene subgraphs, identifying objects, bounding boxes, and their spatial relationships.
- Reinforced Learning for Spatial Grounding: Utilizes a multi-objective reward function that enforces format, count, accuracy, and crucial spatial grounding.
- Comprehensive Spatial Categories: Reasons across nine distinct spatial categories, including relations, reach, size, orientation, instance location, depth, distance, count, and existence.
- High Performance: Demonstrates superior performance compared to GPT-4o on spatial understanding benchmarks, achieved with a relatively small training dataset of 7,587 spatial VQA samples (STVQA-7K).
Good For
- Applications requiring precise 3D spatial understanding from images.
- Tasks involving complex spatial reasoning and object relationship analysis.
- Research into reinforcement learning for multimodal models and spatial AI.