---
title: "Rule-Compliant Visual Spatial Planning for Multimodal Large Language Models: RuleMaze Benchmark and Disentangled Planning"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.20237v1"
permalink: "https://feed.astralor.com/items/itm_dc47b856af3c43484c396efc"
published: "2026-08-20T16:28:28.000Z"
score: 79
channel: "ai"
tags: ["RuleMaze","Interpretable Reasoning","MLLM","Spatial Planning","Rule Following"]
---

# Rule-Compliant Visual Spatial Planning for Multimodal Large Language Models: RuleMaze Benchmark and Disentangled Planning

## AI commentary

> This paper is worth attention: it pushes MLLM evaluation from visual QA to rule-constrained spatial planning, targeting a key weakness of current large models. For researchers in embodied AI, agent planning, or interpretability, RuleMaze and DMP provide strong baselines to compare against.

## At a glance

Multimodal large language models combine linguistic reasoning with visual perception, yet their ability to perform visual spatial planning under explicit or unseen rule constraints remains underexplored. To address this gap, the authors introduce RuleMaze, a controllable benchmark requiring models to navigate mazes while obeying natural-language rules of varying complexity. They also propose Language-Logic-Function Hybridization for automatic rule generation, and Disentangled Multimodal Planning (DMP) that separates perception, execution, and rule verification. Experiments show DMP substantially improves rule compliance and planning success over end-to-end textual baselines, offering transparent intermediate traces for grounded and interpretable spatial planning.

## Key points

- Introduces RuleMaze, a benchmark for evaluating MLLMs' visual spatial planning under natural-language rule constraints.
- Proposes Language-Logic-Function Hybridization to automatically generate rules and convert them into executable logical validators.
- Presents Disentangled Multimodal Planning (DMP) that decouples perception, execution, and rule verification for better generalization and interpretability.
- Experiments show DMP substantially outperforms end-to-end textual planning baselines in rule compliance and planning success.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_dc47b856af3c43484c396efc