Official implementation of "The Pragmatic Persona: Discovering LLM Persona through Bridging Inference" accepted at ICPR 2026.

This repository presents a framework for discovering latent personas in Large Language Models (LLMs) through bridging inference — a cognitive discourse mechanism that captures implicit conceptual relations between utterances. Unlike surface-level approaches, our method reveals how LLMs maintain semantic coherence and express consistent behavioral traits through structured discourse analysis.

Figure 1: Comparison of persona inference with and without bridging inference

• Cognitively Grounded: Based on bridging inference from cognitive discourse theory
• Graph-based Analysis: Constructs semantic knowledge graphs from implicit relations
• Multi-dimensional: Discovers personas across 4 dimensions (Social Role, Personality, Background, Interests)
• Model Agnostic: Works with various LLM scales (1.7B to 80B parameters)

Figure 3: Overview of the Persona-Discovering Agent (PD-Agent) framework

Persona_Bridging/
├── test/
│   ├── test_ds.py              # DeepSeek target-model experiment
│   ├── test_llama.py           # Llama target-model experiment
│   └── test_qwen.py            # Qwen target-model experiment
├── test_qwen_ablation.py       # Qwen ablation study
├── test_llama_ablation.py      # Llama ablation study
├── schema/
│   ├── persona_schema.json         # 4-dimensional persona schema
│   └── bridging_relationships.json # 7 bridging relation types
├── figures/                    # Paper figures
├── requirements.txt            # Python dependencies
└── README.md

Results in the paper were produced on the following configuration:

A CUDA-capable GPU is recommended for running open-source target models (Qwen, Llama, DeepSeek) locally. The PD-Agent reasoning/tool backbone runs through the OpenAI API.

git clone https://github.com/JiSoo-Yang/Persona_Bridging.git
cd Persona_Bridging
pip install -r requirements.txt

Python 3.10+ is recommended.

The experiments use the OpenAI API for the PD-Agent (tool/reasoning) backbone and Hugging Face for loading open-source target models. Create a .env file in the repository root:

# .env
OPENAI_API_KEY=your-openai-key-here
HF_TOKEN=your-huggingface-token-here   # required for gated models (e.g. Llama)

.env is git-ignored and must not be committed. Each script automatically loads it at startup.

Each experiment script is self-contained. The target model and runtime settings are defined at the bottom of each script (no command-line flags). To change them, edit these variables, e.g. in test/test_qwen.py:

TARGET_MODEL = "Qwen/Qwen3-1.7B"   # target LLM to probe
num_questions = 3                  # interview turns
device = "cpu"                     # set to "cuda" to use GPU (recommended)

Then run from the repository root so that .env and the schema/ files are found:

# Qwen target
python test/test_qwen.py

# Llama target
python test/test_llama.py

# DeepSeek target
python test/test_ds.py

At the end of a run, the total execution time is printed:

⏱️  Total execution time: XX.XX seconds

Results are written to an outputs/ directory, including:

• outputs/interview_results_<model>.json — interview dialogue
• outputs/bridging_results.json — extracted bridging relations
• outputs/graph_structure.json — semantic graph data
• outputs/persona_similarity.json — predicted-vs-ground-truth similarity

python test_qwen_ablation.py
python test_llama_ablation.py

Seven canonical bridging relation types based on cognitive discourse theory:

• Mereological: part-of (engine → car), member-of (student → class)
• Frame-related: instrument (knife → cutting), theme (topic → discussion), cause-of (effort → success), in (book → library), temporal (morning → breakfast)

1. Interactive Interview: Generates 3–5 dialogue turns with the target LLM
2. Bridging Extraction: Identifies implicit conceptual relations via few-shot learning
3. Graph Construction: Builds a semantic graph G = (V, E)
4. Centrality Analysis: Computes node importance via degree centrality
5. Persona Inference: Predicts persona attributes from graph structure

The framework was evaluated across 6 reasoning backbones and multiple target LLMs (see Table 3 in the paper).

Cosine similarity with ground-truth personas.