Spatial reasoning remains a major challenge for artificial intelligence, and recent multimodal large language models (MLLMs) are increasingly being evaluated on their ability to understand and manipulate the physical world. Dhruv Anand and Ehsan Shareghi from Monash University, together with their collaborators, introduce Cube Bench, a novel benchmark designed to assess spatial and sequential reasoning through the task of solving a Rubik’s Cube. The benchmark decomposes problem-solving into fundamental skills such as cube face reconstruction, move planning, and error correction, enabling a fine-grained analysis of model performance as puzzle complexity increases. The study reveals a substantial performance gap between leading closed-source and open-weight models and shows that even the most advanced systems struggle as task difficulty grows, highlighting important limitations in current AI capabilities.
The proposed methodology breaks performance into five distinct skills and evaluates models using episodes constructed from a shared set of scrambled cube states to ensure consistency across evaluations. To avoid bias, the authors enforced a near-uniform distribution of correct answers across multiple-choice options, dynamically regenerating test episodes until this condition was met. Strict output-parsing rules were applied, with any deviation from the required one-line format counted as an error, ensuring objective and reproducible scoring. Each test isolates a specific reasoning ability, beginning with cube face reconstruction, which evaluates visual parsing by requiring models to reconstruct cube faces from images, and cross-modal verification, which tests consistency between visual and textual cube representations.
Additional tests probe more advanced reasoning and control capabilities. Optimal move prediction evaluates whether models can select the best action under different input modalities, while reflection-guided re-answering examines whether structured self-correction improves accuracy. Closed-loop step-by-step reasoning assesses whether models can maintain progress over multiple moves, and causal move-effect testing isolates the ability to predict the outcome of an action before executing it. Together, these tasks cover the full “see, evaluate, act, reflect, recover” reasoning loop. Results consistently show that accuracy drops sharply as scramble depth increases, models rarely recover once their solution trajectory diverges, and strong perceptual accuracy does not reliably translate into effective multi-step planning.
The findings demonstrate that closed-source models significantly outperform open-weight models, particularly in complex multi-step control tasks, while open-weight models often perform near chance under challenging conditions. Although simple self-reflection mechanisms offer modest improvements, they can also introduce instability. The authors note that while Cube Bench is limited in scope, focusing on Rubik’s Cubes and relatively shallow scrambles, it is designed to be extensible to longer-horizon tasks. Overall, the benchmark provides a reproducible framework for diagnosing weaknesses in spatial reasoning, action evaluation, and error recovery, and offers a foundation for future research aimed at improving these critical capabilities in artificial intelligence systems.
