Study Reveals Fundamental Limits of Large Language Models in Spatial Reasoning

Open-source large language models exhibit fundamental limits when tasked with solving complex spatial problems, failing to replicate the intuitive navigation of physical environments that humans display, according to a methods study published in Frontiers in Computer Science. The evaluation tested agents powered by five state-of-the-art open-source models to examine the limits of their ability to reason and navigate in complex physical spaces.

The investigation addresses an underexplored area in artificial intelligence research as large language model agents are rapidly deployed as assistive tools in real-world environments. While these systems have demonstrated impressive performance across many reasoning tasks, their spatial reasoning remains underexplored.

To probe these limitations, the research team evaluated the model-powered agents on two distinct, complex spatial problems that humans typically learn to solve with ease: the Tower of Hanoi puzzle and the game of Tetris. Because both tasks are fully describable in text, the experimental design provided the language models with the best conditions to succeed.

In the Tower of Hanoi task, the language model agents initially achieved successful results at lower complexity levels. However, as the puzzle complexity grew, the models failed to discover the general, complexity-independent solution strategies that human solvers typically learn with minimal practice. This strategic limitation culminated in a complete collapse of agent performance as the structural demands of the task escalated.

The models rarely produced legal placements, even in drastically simplified variants of the game.

By evaluating the agents’ chosen trajectories against optimal or acceptable problem-solving paths, the authors documented a fundamental lack of understanding of physical space and physical laws. Across both testing environments, the models routinely failed to adhere to core task rules, diverged into impossible spatial paths, and struggled to track and maintain the state of physical objects over successive steps.

In addition to tracking failures, the agents showed no comprehension of fundamental physical properties, including gravity, rigid bodies, and solid boundaries. Rahman and Sibert concluded that these findings highlight essential differences between human cognition and current language model approaches, pointing to a difficult uphill battle in attempting to scale purely text-based architectures into general problem solvers.

Subject of Research: Spatial reasoning limitations of large language models

Article Title: Lost in space: uncovering fundamental limits of large language models in complex, spatial problem solving

Article References: Rahman, R., & Sibert, C. (2026). Lost in space: uncovering fundamental limits of large language models in complex, spatial problem solving. Frontiers in Computer Science, 8, Article 1932390. https://doi.org/10.3389/fcomp.2026.1932390

Image Credits: AI Generated

DOI: 10.3389/fcomp.2026.1932390

Keywords: large language models, spatial reasoning, artificial intelligence, Tower of Hanoi, Tetris, problem solving, physical laws, AI evaluation, Lost, space, uncovering, fundamental