A Zero-Shot Approach to Decomposing Graphical Rules in Accessibility Regulations

Abstract: Composite graphical rules hinder interpretation by multimodal large language models (MLLMs), and existing annotation-isolation methods require labeled training data. To address this, we propose a training-free approach using the Segment Anything Model 3 (SAM3) to automatically isolate annotations, remove cross-rule visual clutter, and reconstruct individual rules. Using manually decomposed rules as ground truth, we compared our method against prompt-based generative baselines. Results indicate that while baselines produce visually cleaner diagrams, MLLM-as-a-judge evaluations demonstrate our SAM3-based approach more reliably preserves the original regulatory semantics.