Visualization-of-Thought takes inspiration from the human Mind’s Eye to allow LLMs to perform multi-hop spatial reasoning through visualized reasoning traces.

• Employs visual prompts to guide the reasoning process of LLMs.
• Demonstrates enhanced LLM performance in natural and visual navigation, and visual tiling.
• VoT’s performance exceeds existing MLLM capabilities, hinting at potential applications in MLLMs.

VoT’s application in spatial reasoning tasks indicates a directional shift towards AI that can better interpret and navigate spatial environments, mirroring human cognitive processes.