Persona Collapse

Code and data for the paper "The Chameleon’s Limit: Investigating Persona Collapse and Homogenization in Large Language Models".

This codebase supports three experiments:

1. Moral Reasoning -- Evaluate how LLMs respond to moral dilemmas across diverse personas
2. Digital Twin (BFI) -- Replicate human personality assessments (Big Five Inventory) with LLM-simulated personas
3. Self-Introduction -- Collect open-ended self-introductions from LLM-simulated personas

Repository Structure

persona-collapse/
├── persona_sim/             # Shared library
│   ├── consts.py            # Constants (rating scales)
│   ├── prompts.py           # Prompt construction for all experiments
│   ├── providers.py         # Model registries, client factory, Bedrock wrapper
│   ├── experiment.py        # Shared experiment runner (DB, retries, diagnostics)
│   └── metrics.py           # Analysis metrics (Hopkins, PRDC, uniformity)
├── experiments/             # Experiment entry points
│   ├── moral_reasoning.py   # Moral dilemma experiment
│   ├── digital_twin.py      # BFI personality test replication
│   ├── self_introduction.py # Self-introduction collection
│   └── estimate_budget.py   # Token/cost estimation
├── analysis/                        # Post-processing and analysis
│   ├── analysis_pipeline.py         # Full metrics pipeline (BFI, moral, self-intro)
│   ├── check_persona_inconsistency.py  # Generate flagged_personas.pkl
│   ├── extract_csv.py               # Export SQLite results to CSV
│   ├── README.md                    # Analysis folder documentation
│   └── visualizations/              # Figure generation scripts
├── scripts/                         # Shell scripts for running experiments
│   ├── run_moral_reasoning.sh
│   ├── run_digital_twin.sh
│   ├── run_self_introduction.sh
│   ├── run_analysis.sh
│   └── serve_vllm.sh
└── data/                            # Input data files
    ├── sampled_personas_2000.json   # 2000 synthetic persona profiles
    ├── claude-3-5-sonnet_AB_0_with_logprobs.jsonl  # Moral dilemma scenarios
    ├── flagged_personas.pkl         # Pre-computed inconsistent persona IDs
    └── human_reference/
        ├── BFI.json                 # BFI-44 question definitions
        └── wave_1_numbers.csv       # Human survey reference data

Setup

1. Install Dependencies

pip install -r requirements.txt

Core dependencies are openai, numpy, pandas, tqdm, and scikit-learn. Optional dependencies are listed in requirements.txt with comments.

2. Set API Keys

Set the appropriate environment variable for your provider:

# OpenAI
export OPENAI_API_KEY="your-key-here"

# OpenRouter (for models like Kimi, MiniMax, etc.)
export OPENROUTER_API_KEY="your-key-here"

# AWS Bedrock (for Anthropic Claude models)
export AWS_BEARER_TOKEN_BEDROCK="your-key-here"

# Local vLLM (no key needed, but set the base URL)
export VLLM_BASE_URL="http://localhost:8000/v1"

Running Experiments

Supported Providers

Provider	Models	Key
openai	GPT-5, GPT-4.1 family	OPENAI_API_KEY
openrouter	Kimi K2.5, MiniMax M2/M2.5, etc.	OPENROUTER_API_KEY
bedrock	Claude Sonnet/Haiku/Opus	AWS_BEARER_TOKEN_BEDROCK
vllm	Any HuggingFace model via vLLM	VLLM_BASE_URL

Moral Reasoning Experiment

# Step 1: Estimate budget
bash scripts/run_moral_reasoning.sh openrouter kimi-k2.5

# Step 2: Run experiment
bash scripts/run_moral_reasoning.sh openrouter kimi-k2.5 --run

# Or run Python directly:
python -m experiments.moral_reasoning \
    --provider openrouter \
    --model kimi-k2.5 \
    --personas-file data/sampled_personas_2000.json \
    --scenarios-file data/claude-3-5-sonnet_AB_0_with_logprobs.jsonl \
    --use-db 1 \
    --db-path results_moral_kimi-k2.5.db

Digital Twin (BFI) Experiment

# Step 1: Estimate budget
bash scripts/run_digital_twin.sh openrouter minimax-m2 2000

# Step 2: Run experiment
bash scripts/run_digital_twin.sh openrouter minimax-m2 2000 --run

# Or run Python directly:
python -m experiments.digital_twin \
    --provider openrouter \
    --model minimax-m2 \
    --limit 2000 \
    --use-sampled-personas \
    --personas-file data/sampled_personas_2000.json \
    --db-path results_replicate_humans_minimax-m2_bfi.db

Self-Introduction Collection

# OpenRouter (all registered models)
bash scripts/run_self_introduction.sh openrouter

# Specific models
bash scripts/run_self_introduction.sh openrouter kimi-k2.5 minimax-m2

# Local vLLM
bash scripts/run_self_introduction.sh vllm Qwen/Qwen3-4B-Instruct-2507

Using Local Models with vLLM

# Start vLLM server
bash scripts/serve_vllm.sh meta-llama/Llama-3.1-8B-Instruct 8000

# In another terminal, run experiments
export VLLM_BASE_URL="http://localhost:8000/v1"
bash scripts/run_moral_reasoning.sh vllm meta-llama/Llama-3.1-8B-Instruct --run

Advanced Options

Thinking Model Support

For models that produce extended chain-of-thought reasoning (e.g., DeepSeek-R1), enable thinking-model mode. This stores raw outputs and uses an LLM judge to extract numeric ratings:

THINKING_MODEL=true \
THINKING_MAX_TOKENS=16384 \
JUDGE_MODEL=gpt-5-mini \
JUDGE_PROVIDER=openai \
bash scripts/run_moral_reasoning.sh vllm deepseek-r1 --run

Max Tokens

Control the max output tokens for non-thinking mode:

# Moral reasoning (default: 512)
MAX_TOKENS=1024 bash scripts/run_moral_reasoning.sh openrouter minimax-m2 --run

# Digital twin (default: 100)
MAX_TOKENS=200 bash scripts/run_digital_twin.sh openrouter minimax-m2 2000 --run

Flagged Personas

Personas identified as logically inconsistent (e.g., a child persona who is married) can be excluded from API calls to save cost while preserving the full dataset structure. Experiment scripts automatically detect data/flagged_personas.pkl and inject dummy database entries for flagged personas.

To generate the flagged personas file:

python -m analysis.check_persona_inconsistency

This scans data/sampled_personas_2000.json and writes data/flagged_personas.pkl. Use --dry-run to preview without saving.

Resumability

All experiments use SQLite databases for incremental saving. If a run is interrupted, simply re-run the same command -- completed tasks are automatically skipped.

Analysis

Step 1: Export Results to CSV

python analysis/extract_csv.py

This exports all SQLite result databases to CSV files under csv_exports/.

Step 2: Run the Analysis Pipeline

The main analysis pipeline computes all metrics reported in the paper (BFI diagnostics, moral reasoning metrics, self-introduction linguistic analysis):

python -m analysis.analysis_pipeline \
    --data_dir ./csv_exports \
    --selfintro_dir ./self_introduction_results \
    --human_ref ./data/human_reference/wave_1_numbers.csv \
    --output_dir ./results

This produces:

• results/analysis.json -- all BFI + moral metrics per model
• results/selfintro/ -- self-introduction linguistic features, mention rates, ICC, etc.
• results/figures/ -- Coverage-vs-LID and Fidelity-vs-d plots (PDF)

Optional flags:

• --flagged ./data/flagged_personas.pkl -- exclude inconsistent personas
• --skip_selfintro -- skip self-introduction analysis
• --skip_figures -- skip figure generation

Standalone Visualizations

# 3D metric illustrations (coverage, uniformity, complexity)
python analysis/visualizations/metrics_3d.py

# Uniformity figures
python analysis/visualizations/uniformity.py

See analysis/README.md for the full list of analysis scripts and detailed documentation.

Citation

If you use this code or data, please cite our paper:

@article{xiao2026collapse,
  title={The Chameleon’s Limit: Investigating Persona Collapse and Homogenization in Large Language Models},
  author={Yunze Xiao, Vivienne Zhang, Chenghao Yang, Ningshan Ma, Weihao Xuan, Jen-tse Huang},
  journal={arXiv preprint arXiv:[to appear]},
  year={2026},
  note={arXiv ID pending}
}

License

[To be specified]