Researchers have created a benchmark called RuleMaze to test the ability of large language models to perform visual spatial planning under specific rules. The model must navigate mazes while following natural-language instructions that can vary in complexity. To make this process more efficient and scalable, the team developed Language-Logic-Function Hybridization, which automatically generates these rules and translates them into a format that the model can understand. Additionally, they introduced Disentangled Multimodal Planning, which separates perception, execution, and rule verification to improve generalization and provide transparent planning traces. --- Why it matters: This research matters because it addresses a gap in the understanding of how large language models perform visual spatial planning under specific rules. The benchmark and proposed methods can help improve the development of more interpretable and grounded AI systems that can understand and follow complex instructions. Source: https://arxiv.org/abs/2608.20237