# Improved Visual-Spatial Reasoning via R1-Zero-Like Training

Source: [Improved Visual-Spatial Reasoning via R1-Zero-Like Training](<https://www.sota2.com/research/paper/68fa78121e209131cea76a45>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-26T04:31:26.317Z

Authors: Zhenyi Liao, Qingsong Xie, Yanhao Zhang, Zijian Kong, Haonan Lu, Zhenyu Yang, Zhijie Deng

Published: 2025-04-01T15:11:11.000Z

## Abstract

Increasing attention has been placed on improving the reasoning capacities of multi-modal large language models (MLLMs). As the cornerstone for AI agents that function in the physical realm, video-based visual-spatial intelligence (VSI) emerges as one of the most pivotal reasoning capabilities of MLLMs. This work conducts a first, in-depth study on improving the visual-spatial reasoning of MLLMs via R1-Zero-like training. Technically, we first identify that the visual-spatial reasoning capacities of small- to medium-sized Qwen2-VL models cannot be activated via Chain of Thought (CoT) prompts. We then incorporate GRPO training for improved visual-spatial reasoning, using the carefully curated VSI-100k dataset, following DeepSeek-R1-Zero. During the investigation, we identify the necessity to keep the KL penalty (even with a small value) in GRPO. With just 120 GPU hours, our vsGRPO-2B model, fine-tuned from Qwen2-VL-2B, can outperform the base model by 12.1% and surpass GPT-4o. Moreover, our vsGRPO-7B model, fine-tuned from Qwen2-VL-7B, achieves performance comparable to that of the best open-source model LLaVA-NeXT-Video-72B. Additionally, we compare vsGRPO to supervised fine-tuning and direct preference optimization baselines and observe strong performance superiority. The code and dataset will be available soon.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Target Viewpoint Reproduction on TVRBench (test)](<https://www.sota2.com/research/sota/target-viewpoint-reproduction-on-tvrbench-test.md>): Target Viewpoint Reproduction; TVRBench (test); Overall Performance: 26.2.

- [Spatial Reasoning on VSI-Bench (test)](<https://www.sota2.com/research/sota/spatial-reasoning-on-vsi-bench-test.md>): Spatial Reasoning; VSI-Bench (test); Avg Score: 40.7.

- [Spatial Reasoning on VSI-Bench](<https://www.sota2.com/research/sota/spatial-reasoning-on-vsi-bench.md>): Spatial Reasoning; VSI-Bench; R.Dr.: 35.6.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2504.00883>)

[PDF](<https://arxiv.org/pdf/2504.00883.pdf>)