Current Vision-Language Models (VLMs) struggle with fine-grained spatial reasoning, particularly when multi-step logic and precise spatial alignment are required. In this work, we introduce SpatialReasoner-R1, a vision-language reasoning model designed to address these limitations. To construct high-quality supervision for spatial reasoning, we design a Multi-Model Monte Carlo Tree Search (M3CTS) method that generates diverse, logically consistent Long Chain-of-Thought (LongCoT) reasoning trajectories. In addition, we propose fine-grained Direct Preference Optimization (fDPO), which introduces segment-specific preference granularity for descriptive grounding and logical reasoning, guided by a spatial reward mechanism that evaluates candidate responses based on visual consistency, spatial grounding, and logical coherence. Experimental results demonstrate that fDPO achieves an average improvement of 4.1% over standard DPO across spatial quality tasks, and a 9.0% gain in spatial quantity tasks. SpatialReasoner-R1, trained with fDPO, sets a new SoTA on SpatialRGPT-Bench, outperforming the strongest baseline by 9.8% in average accuracy, while maintaining competitive performance on general vision-language tasks.
Method Overview including SpatialReasoner-R1 model architecture and training pipeline. Training pipeline consisting of three stages: (1) generating reasoning paths using M3CTS; (2) constructing fine-grained preference pairs via reward-based selection; (3) training with fine-grained DPO (fDPO) to optimize descriptive and logical reasoning separately.
Fine-Grained Spatial Rewards. Candidate reasoning paths are decomposed into three aspects, descriptive, spatial, and reasoning, scored separately; the higher value in each row is marked by ✔ and the lower by ✖.
We conduct comprehensive evaluation on spatial reasoning tasks to demonstrate the effectiveness of our approach.
Spatial reasoning success rates (↑) on SpatialRGPT-Bench. "/" indicates that the model refuses to provide a response for that metric. SpatialReasoner-R1 8B, trained with fDPO, establishes a new SoTA in spatial reasoning.
SpatialReasoner-R1 demonstrates improved spatial reasoning across various scenarios and object types:
@misc{shen2025finegrainedpreferenceoptimizationimproves,
title={Fine-Grained Preference Optimization Improves Spatial Reasoning in VLMs},
author={Yifan Shen and Yuanzhe Liu and Jingyuan Zhu and Xu Cao and Xiaofeng Zhang and Yixiao He and Wenming Ye and James Matthew Rehg and Ismini Lourentzou},
year={2025},
eprint={2506.21656},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2506.21656},
}