Discovering the Latent Persona of Large Language Models via Bridging Inference

Large Language Models (LLMs) reveal their inherent and distinctive personas through dialogue. However, existing approaches often rely on surface-level lexical and stylistic cues, making it difficult to capture the deeper contextual meanings required for accurate persona discovery. To address this limitation, we propose a novel analytical framework that models implicit conceptual relations in dialogue as bridging inferences and represents them as knowledge graphs. This structure enables a systematic understanding of the hidden semantic connections underlying discourse, providing deeper insights into how LLMs consistently express latent traits throughout conversation. Experimental results show that bridging-inference graphs capture stronger semantic coherence than existing methods, allowing consistent identification of the target LLM's assigned persona characteristics. To our knowledge, this study presents the first systematic attempt to probe, extract, and visualize latent LLM personas through the lens of Cognitive Discourse Theory, thereby bridging computational linguistics, cognitive semantics, and persona reasoning in large language models.