The Pragmatic Persona: Discovering LLM Persona through Bridging Inference

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Linguistics · Depth: Expert, extended

Summary

The PD-Agent framework introduces a novel approach to discovering latent personas in Large Language Models (LLMs) by analyzing "bridging inference," which are implicit conceptual relations connecting dialogue utterances. Unlike methods relying on surface-level lexical cues, PD-Agent models these relations as structured knowledge graphs, capturing deeper discourse coherence. The framework operationalizes seven bridging relation types and employs a three-phase process: interactive interviewing, few-shot exemplar-based relation extraction, and graph-based reasoning using normalized degree centrality. Experiments across six reasoning backbones (including GPT-4o, Claude 3.5 Sonnet, and o1-mini) and target LLMs up to 80B parameters (e.g., Llama-3.1-70B, Qwen3-80B) demonstrated PD-Agent's superior performance, achieving 0.90 to 0.98 cosine similarity with ground-truth personas, significantly outperforming frequency-based baselines by up to +0.15. The framework's effectiveness scales positively with target model size, providing a robust diagnostic tool.

Key takeaway

For NLP Engineers and AI Scientists focused on understanding or steering LLM behavior, you should move beyond surface-level lexical analysis for persona discovery. Instead, consider implementing discourse-level relational reasoning, such as bridging inference, to uncover deeper, more stable persona traits. This approach, especially effective with larger models, provides a robust diagnostic tool for evaluating model consistency and can inform more precise persona alignment strategies, even if "background" attributes might require more nuanced cue interpretation.

Key insights

LLM personas are deeply encoded in discourse coherence via implicit conceptual links, not just surface-level lexical patterns.

Principles

• Bridging inference reveals latent semantic links in dialogue.
• Persona traits are consistently encoded in discourse structure.
• Graph-based reasoning enhances persona identification stability.

Method

The PD-Agent conducts adaptive interviews, extracts seven types of bridging relations using few-shot exemplars, then constructs a directed semantic graph. Persona attributes are inferred by analyzing node centrality and relation type distribution.

In practice

• Model LLM dialogue as knowledge graphs for deeper analysis.
• Use normalized degree centrality to identify key concepts.
• Evaluate persona consistency beyond lexical features.

Topics

• Large Language Models
• Persona Discovery
• Bridging Inference
• Cognitive Discourse Theory
• Knowledge Graphs
• Dialogue Systems

Code references

• JiSoo-Yang/Persona_Bridging

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Adaptive Interviewing for Persona Simulation in LLMs: Evidence-Grounded Reasoning Improves Decision Alignment — Adaptive interviewing improves LLM decision simulation by selectively grounding predictions in user-specific, follow-up derived evidence.
• Steering at the Source: Style Modulation Heads for Robust Persona Control — Targeting specific "Style Modulation Heads" in LLMs enables robust persona control with minimal coherency degradation.
• Unveiling the Limits of Large Language Models in Inferring Pragmatic Meaning from Non-Verbal Responses — LLMs significantly struggle to infer pragmatic meaning from non-verbal communication, unlike their performance with verbal cues.
• Research on LLM alignment as latent discourse-level regimes vs. token-level filtering? — LLM alignment may stem from global latent space geometry and discourse-level regimes, not just token-level filtering.
• The persona selection model — LLMs simulate diverse personas during pre-training, with post-training refining a specific "Assistant" character.
• NICE: A Theory-Grounded Diagnostic Benchmark for Social Intelligence of LLMs — NICE provides a theory-grounded diagnostic benchmark to identify specific social intelligence weaknesses in LLMs, crucial for their application in social contexts.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.