The Pragmatic Persona: Discovering LLM Persona through Bridging Inference

The article introduces a novel analytical framework for discovering personas in Large Language Models (LLMs) through "bridging inference," which utilizes structured knowledge graphs to capture deeper discourse-level structures rather than relying on surface-level lexical cues. Experimental results demonstrate that this approach, tested on various models up to 80 billion parameters, significantly enhances semantic coherence and stability in persona identification compared to traditional frequency or style-based methods. This framework has implications for practitioners by providing a systematic method for probing and visualizing LLM personas, integrating insights from cognitive discourse theory with computational linguistics.