RuleMaze Benchmark for Rule-Compliant Visual Spatial Planning in MLLMs

RuleMaze introduces a benchmark for evaluating how multimodal large language models execute spatial planning under natural-language rule constraints. It utilizes a Language-Logic-Function Hybridization method to automatically generate rules and executable validators for scalable testing.