PDOS: Widening the Safety Overton Window Through Persona-Driven Generation
Abstract
AbstractSafety in large language models is often framed as requiring a single correct response, even in safety-sensitive contexts where multiple perspectives may exist. Existing approaches largely ignore Overton pluralism (inclusion of multiple reasonable viewpoints for a single query) and reduce complex context to monolithic judgments that fail to capture the range of reasonable viewpoints. Moreover, existing frameworks are not grounded in safety settings, which limits their ability to capture the diversity of perspectives required in safety-critical situations. Existing safety datasets do not provide Overton-style annotations, limiting their ability to represent multiple harm perspectives within the same safety-critical situation. To address this, we construct **SOS**, the safety-focused Overton dataset of 1,770 real-world safety situations annotated across 14 harm taxonomy. We further propose **PDOS**, a lightweight, training-free framework that retrieves personas from a synthetic pool and integrates their perspectives to generate unified Overton safety responses. PDOS achieves the highest average NLI coverage of 0.4226 among several baselines, and outperforms the strongest baseline by 13.3%. Our findings show that Overton-based safety pluralism requires adaptable, context-aware approaches, providing insights into how these methods can better capture diverse perspectives.