Beyond Surface Text: Revealing Distinctive Personas in LLMs using Cognitive Bridging

Large Language Models (LLMs) express distinctive personas through dialogue, yet most existing approaches rely on surface-level lexical or stylistic cues, making it difficult to uncover the deeper contextual meanings behind their behavior. To address this limitation, we propose a cognitively grounded framework that models dialogue understanding in LLMs through bridging inference—the human process of linking implicitly related entities and events using world knowledge. Our approach introduces a two-agent system consisting of a Persona Discovering Agent (PD-Agent) and a Target LLM (e.g., Qwen3-1.7B, LLaMA3.1-8B, Gemini-2.5-Flash). The framework proceeds through three stages: (1) Interview Stage – the PD-Agent conducts adaptive interviews with the Target LLM, which has been assigned a hidden persona derived from four structured dimensions: social role, personality, background, and interests; (2) Bridging Inference Extraction – linguistic definitions from Irmer’s (2011) seven-category schema (e.g., part–whole, instrument–event, cause–effect) are used to identify implicit conceptual relations across utterances; (3) Graph Construction and Persona Prediction – extracted relations are represented as a directed graph, where node importance and edge density reveal dominant reasoning patterns that inform persona inference. Quantitative evaluation measured cosine similarity between predicted and ground-truth personas. Our framework achieved an average similarity of 0.94, outperforming surface-only (0.80) and frequency-aware (0.87) baselines across all target models. These results demonstrate that bridging-based discourse reasoning captures deeper semantic and social coherence than lexical or statistical methods, enabling interpretable persona discovery grounded in cognitive discourse theory. By uniting cognitive linguistics and computational modeling, this work highlights how LLMs construct and maintain coherent social identities through implicit reasoning—offering new insights into the cognitive foundations of AI communication.