Mind's Eye of LLMs: Visualization-of-Thought Elicits Spatial Reasoning in Large Language Models

Authors: Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, Furu Wei

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrated that Vo T significantly enhances the spatial reasoning abilities of LLMs.
Researcher Affiliation Collaboration Wenshan Wu Shaoguang Mao Yadong Zhang , , Yan Xia Li Dong Lei Cui Furu Wei Microsoft Research East China Normal University
Pseudocode Yes Algorithm 1: Navigation Map Generation
Open Source Code Yes Please find the dataset and codes in our project page.
Open Datasets Yes The data and code associated with this study is publicly available and the link is provided in the paper.
Dataset Splits Yes Visual Navigation We generate 496 navigation maps and 2520 QA instances in total, covering various map sizes, up to 7x9 and 9x7. The data distribution is provided in Table 4 in appendix.
Hardware Specification No API settings are temperature 0 as greedy decoding and top p 1, with model versions of 1106-preview and vision-preview.
Software Dependencies Yes Specifically, we adopt GPT-4 [OA+23] and GPT-4 Vision [Ope23] via Azure Open AI API as they re state of the art LLM and multimodal model respectively. API settings are temperature 0 as greedy decoding and top p 1, with model versions of 1106-preview and vision-preview.
Experiment Setup Yes API settings are temperature 0 as greedy decoding and top p 1, with model versions of 1106-preview and vision-preview. For all experiments we adopt zero-shot prompting.