RuleMaze: A Benchmark for Rule-Compliant Visual Spatial Planning in Multimodal Large Models
This paper focuses on the visual spatial planning capability of multimodal large language models (MLLMs) under explicit or unseen rule constraints, noting that this ability remains underexplored. The authors propose RuleMaze, a controllable benchmark that requires models to navigate a maze while following rules of varying natural-language complexity, thereby isolating three core capabilities: perception, rule interpretation, and constrained action planning. To systematically construct rules, the authors introduce a Language-Logic-Function Hybridization approach that automatically generates natural-language rules and translates them into logical representations and executable verifiers, eliminating manual rule engineering. To improve rule-following and generalization, the authors further propose Decoupled Multimodal Planning (DMP), which separates perception, execution, and rule verification through interpretable reasoning primitives, enabling systematic generalization to more complex and unseen rules while providing transparent intermediate planning trajectories. Experiments show that DMP significantly outperforms end-to-end text planning baselines in rule compliance and planning success rate. RuleMaze provides a principled benchmark for studying grounded and interpretable rule-driven spatial planning.
Background and Context
Multimodal large language models combine language reasoning with visual perception, yet their ability to perform spatial planning under explicit or previously unseen rule constraints remains underexplored. This scenario forces a model to accomplish three things simultaneously: understand a spatial layout, interpret natural-language rules, and plan a sequence of legal actions that respect them. To address this gap, researchers introduce RuleMaze, a controllable maze-navigation benchmark that isolates rule-compliant spatial planning from ordinary visual tasks.
RuleMaze requires models to navigate mazes while following rules of varying natural-language complexity. By systematically varying rule complexity, the benchmark separates three core capabilities: perception, rule interpretation, and constrained action planning. This design converts the vague question of whether a model "follows rules" into a quantifiable, decomposable, and reproducible scientific problem.
The work contributes both a principled evaluation benchmark and a corresponding methodological framework. Rather than stopping at task-pass rates, it encourages researchers to examine the structure and explainability of the planning process itself, opening a clear entry point for studying grounded, rule-driven spatial reasoning.
Deep Analysis
The paper's first technical contribution is Language-Logic-Function Hybridization, a method for systematically and scalably constructing rules. It automatically generates natural-language rules and translates them into logical representations and executable verifiers, eliminating manual rule engineering. This hybrid design lets rules be expressed in human language while being strictly validated by machine logic, preserving both scalability and consistency.
To improve rule-following and generalization, the authors propose Decoupled Multimodal Planning (DMP). Using interpretable reasoning primitives, DMP explicitly separates perception, execution, and rule verification. Instead of folding all three tasks into a single end-to-end text stream, the model processes spatial perception, action execution, and rule checking independently.
This structured decomposition lets DMP generalize systematically to more complex and unseen rules while providing transparent intermediate planning trajectories. Every decision step becomes traceable and explainable. Compared with end-to-end approaches that compress perception, reasoning, and action into one text generation, DMP offers clear advantages in controllability and interpretability.
Industry Impact
On the RuleMaze benchmark, the authors systematically evaluate multiple methods, focusing on comparing end-to-end text planning baselines against DMP. Results show that DMP significantly outperforms the baselines on both rule compliance and planning success rate, validating that the decoupled design genuinely helps rule-following and generalization.
Ablation studies point to the same conclusion: separating perception, execution, and rule verification produces more stable and more explainable planning than modeling them implicitly together. RuleMaze's value lies in treating rule complexity as a controllable variable, allowing researchers to observe how models degrade and adapt under more complex, unseen rules.
For the open-source community, the paper releases code, lowering the barrier to reproduction and extension. For industrial deployment, tasks requiring strict rule adherence—robot navigation, automation, game AI, and embodied intelligence—depend on exactly this kind of verifiable, explainable planning. DMP's transparent intermediate trajectories offer auditable decision evidence for such high-risk scenarios.
Outlook
RuleMaze reframes rule-compliant planning from a vague capability description into a decomposable, quantifiable scientific question, encouraging further exploration of how perception, reasoning, and verification interact. The results suggest that improving planning capability depends not only on model scale but also on the sensible decomposition of task structure.
Overall, the work signals to the industry that evaluating multimodal large models should move beyond task-pass rates and dig into the structure and explainability of planning processes. This shift pushes models from merely "solving problems" toward acting reliably and according to rules.
RuleMaze thus provides a principled benchmark for studying grounded and interpretable rule-driven spatial planning, setting the stage for future research on trustworthy, verifiable multimodal decision-making.