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

Large language models (LLMs) have exhibited impressive performance in language comprehension and various reasoning tasks. However, their abilities in spatial reasoning, a crucial aspect of human cognition, remain relatively unexplored. Human possess a remarkable ability to create mental images of unseen objects and actions through a process known as the Mind's Eye, enabling the imagination of the unseen world. Inspired by this cognitive capacity, we propose Visualization-of-Thought (VoT) prompting. VoT aims to elicit spatial reasoning of LLMs by visualizing their reasoning traces, thereby guiding subsequent reasoning steps. We employed VoT for multi-hop spatial reasoning tasks, including natural language navigation, visual navigation, and visual tiling in 2D grid worlds. Experimental results demonstrated that VoT significantly enhances the spatial reasoning abilities of LLMs. Notably, VoT outperformed existing multimodal large language models (MLLMs) in these tasks. While VoT works surprisingly well on LLMs, the ability to generate mental images to facilitate spatial reasoning resembles the mind's eye process, suggesting its potential viability in MLLMs. Please find the dataset and codes at https://microsoft.github.io/visualization-of-thought

Paper

Similar papers

Peer review

Reviewer Af548/10 · confidence 4/52024-07-05

Summary

The paper explores the spatial reasoning capabilities of LLMs. To this end, the paper introduces novel testbeds including multiple tasks and, importantly, a novel approach called Visualization-of-Thought (VoT). This method aims to enhance the spatial reasoning capabilities of LLMs by visualizing their reasoning processes, also inspired by chain of thought CoT reasoning and human cognitive abilities. Evaluating on the mentioned testbed the authors show that VoT significantly improves the spatial reasoning performance of LLMs, outperforming e.g. CoT reasoning.

Strengths

- The introduction of Visualization-of-Thought (VoT) represents a novel problem and significant advancement in eliciting spatial reasoning in LLMs, a relatively unexplored area. - The paper is well-structured, balancing necessary details in the main text with extensive supplementary information in the appendix. The problem, tasks as well as the VoT approach are well motivated and described. - The paper presents well-conducted and detailed experiments across multiple tasks, providing robust evidence of VoT’s effectiveness. Further, the experimental setup is well described. Additionally, the consideration of openly available models (LLaMA), enhances the study’s reproducibility. - The authors provide a thorough discussion of the results and the limitations providing valuable insights for future work.

Weaknesses

While novel, the introduced and investigated tasks to evaluate spatial reasoning capabilities of LLMs may still be relatively simple compared to real-world spatial reasoning challenges. However, this is negligible. The paper’s contribution provides significant novelty and lays a robust foundation for future exploration and advances of spatial reasoning in large language models.

Questions

- How might VoT be adapted or scaled to work with 3D spatial reasoning tasks or other modalities?

Rating

8

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

The authors have adequately addressed the limitations of their work.

Reviewer yVwi5/10 · confidence 3/52024-07-08

Summary

The paper proposes a new prompting method, "Visualization of Thought" (VoT) prompting, to enable LLMs to perform better on spatial reasoning tasks. In VoT, the LLM is prompted with instructions on performing a multi-step spatial reasoning task, followed by the text "Visualize the state after each reasoning step". The authors evaluate this prompting method on three types of spatial reasoning tasks, and show that it substantially improves performance on these tasks as compared with Chain-of-Thought prompting, as well as with other variations on prompting.

Strengths

VoT prompting is an interesting and (I think) novel approach to elicit better spatial reasoning in LLMs, and may be useful in multimodal systems as well (this is future work). The evaluated tasks are chosen to capture important aspects of spatial reasoning abilities. The paper reports on systematic experiments that evaluate these abilities on the chosen tasks.

Weaknesses

There are three categories of weaknesses which I believe can be improved on by the authors. 1. Some aspects of the paper are unclear and need more explanation. -- The paper says "This method augments LLMs with a visuospatial sketchpad". It's unclear what is meant here by visuospatial sketchpad. This sentence makes it sound like there is a new architectural feature added to the LLM, but my understanding from the rest of the paper is that VoT just asks the LLM to generate a series of "visualizations" in its output. Please make it clearer in the description what exactly you mean. -- Figure 4 gives examples of VoT prompting. But it wasn't clear to me whether these prompts include actual images (e.g., the pictures of barriers and people in the navigation task, or the pictures of colored and empty boxes in the visual tiling task) or whether these images in the figures were actually translated into some kind of text encoding for the LLM. Please make clear exactly what is in the prompt---images or text---and if text, what exactly is the text encoding of the pictures you are showing? -- Equation 5 defines v_i as "visualizations". What exactly are these, in terms of LLMs? Are they text or image outputs from the LLMs? -- In several figures (e.g., Figure 4, 6, 7) prompts and responses are shown but it needs to be made very clear which part is the prompt and which part the response from the LLM. These were not clear enough for me. -- In Figure 8, again, what is part of the prompt and what is part of the response? -- Figure 10: The caption only says "mental images for state tracking in visual navigation tasks." Please say more in the caption about what this figure is showing. Was "Use round pin" part of the prompt, or is this just describing what the LLM did? What are you trying to show in this figure? In general, please make the figure captions more informative. -- "Specifically, we adopt GPT-4 and GPT-4 Vision" -- there are several versions of each of these LLMs on the Azure platform -- please specify which ones you used for these experiments. -- "We adopted sub-string matching as f_correct to determine correctness." This didn't really explain how you determined correctness. Please give more details (enough for someone else to replicate) and explain why this sub-string matching method is reliable for assessing correctness. 2. Scaling: The tasks used were relatively simple: 3x3 maps for navigation, 5x4 rectangle for tiling. Did you do any experiments with scaling up the size of these? How does this method's spatial reasoning abilities scale as the size is increased? 3. Anthropomorphism: The paper is filled with anthropomorphic language, which makes it difficult to understand. For example: -- "LLMs possess the ability to create and manipulate mental images in the mind's eye." Using terms like "mental images" and "the mind's eye" assumes that the LLM is doing something analogous to the brain, which is not necessarily true! -- "visual state tracking is implemented by generating the visualization v_i as a mental image of the internal state si after each reasoning step z". A mental image is something a human has. Using that term for LLMs is unhelpful anthropomorphizing. -- The paper calls the intermediate outputs of LLMs "thoughts" (e.g., "this process by adding a visuospatial sketchpad to each thought z_i") -- again, an unhelpful and confusing anthropomorphism. -- "We demonstrate LLMs’ zero-shot capability of visualizing their precepts at abstract level". Visualizing is a cognitive ability of brains. Saying that LLMs "visualize" is an incorrect anthropomorphism. -- ""this emergent capability to create mental images to enhance spatial reasoning resembles the mind’s eye process, suggesting its promise in MLLMs." Same comments as above.

Questions

-- I was confused by Figure 8b. The left side shows the house moving -- is that correct? Also, it doesn't do what the instructions say: e.g., "move down to the end of the continuous road" -- after the move the house only moved down one square, not two. Is this considered a correct solution? The right side shows a person moving. Where did the notion of moving a person versus moving a house come from?

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

The authors have adequately addressed the limitations.

Reviewer us5D5/10 · confidence 4/52024-07-12

Summary

This paper focuses on enhancing the spatial reasoning capabilities of Large Language Models (LLMs) by introducing Visualization-of-Thought (VoT) prompting. Inspired by the human cognitive ability to visualize unseen objects - a process known as the Mind’s Eye - VoT visualizes reasoning processes to guide LLMs in tasks like natural language navigation and visual navigation in 2D grid worlds.

Strengths

1. The introduction of Visualization-of-Thought (VoT) prompting represents a new method that mimics human cognitive processes, specifically the ability to visualize mental images for better spatial reasoning. 2. To investigate the effectiveness of the proposed VoT, this paper considers three spatial-related tasks, including natural language navigation, visual navigation, and visual tiling.

Weaknesses

**Method Some descriptions are unclear. 1. The method for obtaining the visuospatial sketchpad $v$ remains unclear. This paper describes $v$ as representing a visualization sequence in text form, but it does not specify how visual information (maybe images) is converted into text. 2. If visual information is transformed into text form, how does GPT-4V use it? Is it used as a textual representation of the visual image as well? **Experiment 1. This paper also introduces the concept of state $s$. What is the performance comparison when this state is included versus when it is excluded? 2. In Table 1, for the visual navigation and visual tiling tasks, why does GPT-4V CoT perform the poorest, even yielding results lower than GPT-4 without visual information? 3. It seems that Section 5.2 aims to discuss “how visualization enhance final answers”, yet the connection between the results in Table 2 and this discussion remains unclear. It would be beneficial to include more detailed descriptions or results related to this. 4. In Section 5.3, Table 3 shows that Llama3-8B with VoT underperforms in most results compared to its counterpart with CoT. Does this suggest that the proposed VoT does not benefit less powerful language models?

Questions

Please refer to weaknesses.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

This paper provides the limitations.

Reviewer MbkM6/10 · confidence 5/52024-07-13

Summary

The paper presents a novel approach called Visualization-of-Thought (VoT) prompting, aimed at enhancing the spatial reasoning abilities of Large Language Models (LLMs). Inspired by the human cognitive process known as the “Mind’s Eye,” the authors propose a method where LLMs visualize their reasoning steps to guide subsequent reasoning. The paper evaluates VoT across three spatial reasoning tasks: natural language navigation, visual navigation, and visual tiling. Experimental results demonstrate significant performance improvements with VoT over conventional prompting methods and existing multimodal large language models (MLLMs).

Strengths

Strengths 1. **Innovative Direction**: The concept of eliciting spatial reasoning through visualization is novel and well-motivated. Drawing parallels with human cognition adds a strong cognitive science foundation to the work. 2. **Comprehensive Evaluation and Improved Performance**: The paper evaluates VoT across multiple tasks and benchmarks, providing a thorough analysis of its effectiveness. The results show that VoT significantly enhances spatial reasoning abilities in LLMs, outperforming existing methods. Meawhile, this paper includes detailed qualitative and quantitative analyses, shedding light on the mechanisms behind VoT’s success and its limitations.

Weaknesses

Weaknesses 1. **Limited Task Diversity**: While the selected tasks cover various aspects of spatial reasoning, the evaluation could be broadened to include more complex and varied real-world scenarios. In the main paper, we even didnt see more commonly adopted benchmark performance comparison. Include more varied and complex spatial reasoning tasks to better evaluate the generalizability and robustness of VoT. 2. **Generalizability**: The paper primarily focuses on GPT-4 and GPT-4V. It would be beneficial to see how VoT performs across a wider range of models, particularly less advanced ones. It would be great to see more evaluations of VoT on a wider range of models, including less advanced ones, to assess its applicability across different LLMs. 3. **Prompt Sensitivity**: The effectiveness of VoT appears to be sensitive to the specific wording of prompts, which may limit its robustness in practical applications. Investigate ways to reduce the sensitivity of VoT to specific prompt wording, making it more robust for practical use.

Questions

See above Weakness part.

Rating

6

Confidence

5

Soundness

4

Presentation

4

Contribution

3

Limitations

See above Weakness part.

Reviewer yVwi2024-08-09

Thank you for the rebuttal

Thank you for the thoughtful rebuttal. This cleared up some aspects of the method and results for me -- I hope you will include all of these clarifications in a revised version of the paper.

Authorsrebuttal2024-08-11

We sincerely appreciate your feedback on our rebuttal. We're glad to hear that our responses helped clarify aspects of our methodology and findings. We will certainly incorporate all the clarifications provided in the rebuttal into a revised version of the paper to ensure it is as clear and comprehensive as possible for all readers. Your comments have been instrumental in improving the quality and clarity of our manuscript. If you have any remaining concerns, please let us know, and we will be more than willing to address them.

Reviewer Af542024-08-12

Thank you for the rebuttal. I have no further questions. After carefully considering the other reviews and the authors’ responses, I would keep my initial score.

Reviewer us5D2024-08-13

Thanks for your detailed response. It has addressed my concerns, I will raise the score to reflect this.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC