Rule-Compliant Visual Spatial Planning for Multimodal Large Language Models
2026-08-20 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors created RuleMaze, a test where multimodal large language models (MLLMs) have to navigate mazes while following natural language rules that can be new or tricky. They made a system to automatically create and check these rules, so they don't have to write them by hand. To help models do better, the authors designed Disentangled Multimodal Planning (DMP), which breaks down the task into parts like seeing, planning, and checking rules separately. Their experiments showed that DMP helps models follow rules more accurately and handle new challenges better than usual methods. RuleMaze provides a clear way to study how well models can do spatial planning using language and vision together.
Multimodal Large Language ModelsSpatial PlanningNatural Language RulesRuleMaze BenchmarkLanguage-Logic-Function HybridizationDisentangled Multimodal PlanningRule ComplianceInterpretable ReasoningMaze NavigationControlled Benchmark
Authors
Yu Chen, Ting Lei, Yaoyi Li, Jia Cai, Zhecen Wu, Yang Liu
Abstract
Multimodal large language models (MLLMs) combine linguistic reasoning with visual perception, yet their ability to perform visual spatial planning under explicit or previously unseen rule constraints remains underexplored. This setting requires models to jointly understand spatial layouts, interpret natural-language rules, and plan valid actions accordingly. To address this gap, we introduce RuleMaze, a controllable benchmark in which MLLMs must navigate mazes while obeying natural-language rules of varying complexity. RuleMaze isolates rule-compliant spatial planning by requiring accurate perception, rule interpretation, and constrained action planning. To enable scalable and systematic rule construction, we propose Language-Logic-Function Hybridization, which automatically generates natural-language rules and translates them into logical representations and executable validators, eliminating manual rule engineering. To improve rule following and generalization, we introduce Disentangled Multimodal Planning (DMP), which separates perception, execution, and rule verification through interpretable reasoning primitives. By disentangling these components, DMP facilitates systematic generalization to more complex and previously unseen rules, while providing transparent intermediate planning traces. Experiments demonstrate that DMP substantially improves rule compliance and planning success compared to end-to-end textual planning baselines. Overall, RuleMaze establishes a principled benchmark for studying grounded and interpretable rule-based spatial planning in MLLMs. Code is available at https://github.com/oceanflowlab/RuleMaze.