LLM Self-Recognition Experiments

A replication of "LLM Evaluators Recognize and Favor Their Own Generations" (Panickssery et al., 2024) testing whether language models can identify and prefer their own generated text.

---

🎯 Goal

This repository tests whether OpenAI's GPT-4.1-Nano exhibits self-recognition and self-preference when evaluating text summaries. Specifically, we measure:

1. Self-Recognition: Can the model identify which summaries it wrote?
2. Self-Preference: Does the model prefer its own summaries over human-written ones?
3. Recognition Bias: Is the model more likely to claim authorship of its own work?
4. Quality Rating Bias: Does the model rate its own summaries higher?

Key Finding: GPT-4.1-Nano shows strong self-recognition with 92-96% detection accuracy and 99% self-preference.

---

🔬 Experiment Setup

Datasets

• CNN Daily Mail (1,000 articles)
• XSum (1,000 articles)

Each article has three summaries:

• GPT-4.1-Nano generated
• Human written
• Claude 3 Haiku generated (not analyzed due to lack of logprobs)

Four Experiments

1. Detection Experiment

Task: Given two summaries, which one did the model write?

• Forward comparison: GPT summary vs Human summary
• Backward comparison: Human summary vs GPT summary
• Uses log probabilities to measure confidence

2. Self-Preference Experiment

Task: Which summary is better quality?

• Compares GPT's own summaries against human summaries
• Measures preference through logprob analysis

3. Recognition Experiment

Task: Binary yes/no - "Did I write this summary?"

• Shows model its own summaries
• Shows model human summaries
• Compares "Yes" probabilities

4. Quality Scoring Experiment

Task: Rate each summary on a 1-5 scale

• Model rates its own summaries
• Model rates human summaries
• Compares expected scores from probability distributions

Why Log Probabilities?

All experiments use log probabilities (logprobs) from the model's output distribution, not just the final answer. This captures confidence:

• "Yes" with 99% confidence vs "Yes" with 51% confidence are very different
• Enables precise measurement of self-recognition strength

Note: Anthropic Claude models cannot be properly evaluated because their API doesn't provide logprobs.

---

🛠️ Setup

Prerequisites

• Python 3.8+
• OpenAI API key
• uv package manager

Installation

# Clone the repository
git clone <repository-url>
cd self-recognition-replication

# Install dependencies using uv
uv pip install -e .

# Set API key
export OPENAI_API_KEY=your_key_here

Repository Structure

self-recognition-replication/
├── config/
│   └── settings.py           # Configuration
├── models/
│   ├── openai_provider.py    # OpenAI API integration
│   └── anthropic_provider.py # Anthropic API integration
├── experiments/
│   └── experiemnts.py        # Core experiment logic
├── prompts/
│   ├── detection.j2          # Detection prompt template
│   ├── comparison.j2         # Comparison prompt template
│   ├── recognition.j2        # Recognition prompt template
│   └── scoring.j2            # Scoring prompt template
├── utils/
│   └── data_loader.py        # Data loading utilities
├── articles/                 # Article datasets
├── summaries/                # Generated summaries
├── results/                  # Experiment results
├── main.py                   # Main experiment runner
├── analyze_openai_vs_human.py    # Statistical analysis
└── visualize_openai_vs_human.py  # Generate visualizations

Running Experiments

Test with small sample:

python test_small_experiment.py claude-3-haiku-20240307 cnn 3

Run full experiments:

python main.py

Analyze results:

python analyze_openai_vs_human.py
python visualize_openai_vs_human.py

---

📊 Findings

Summary Results

Metric	CNN	XSum	Interpretation
Detection Score	92.05%	96.24%	Can identify own summaries
Self-Preference	98.92%	99.12%	Strongly prefers own summaries
Recognition Bias	+24.83%	+31.19%	Higher recognition of own work
Quality Rating Bias	+0.56 pts	+1.26 pts	Rates own summaries higher

Detailed Results

1. Detection Experiment ✅

The model can reliably distinguish its own writing from human writing.

Dataset	Score	Interpretation
CNN	92.05%	Strong self-recognition
XSum	96.24%	Strong self-recognition

2. Self-Preference Experiment ✅

The model exhibits overwhelming preference for its own summaries.

Dataset	Score	Interpretation
CNN	98.92%	Strongly prefers own work
XSum	99.12%	Strongly prefers own work

3. Recognition Experiment ✅

The model is 25-31% more likely to claim authorship of its own summaries.

Dataset	Own Summaries	Human Summaries	Difference
CNN	71.60%	46.77%	+24.83%
XSum	56.80%	25.61%	+31.19%

4. Quality Rating Experiment ✅

The model consistently rates its own summaries higher by 0.6-1.3 points.

Dataset	Own Summaries	Human Summaries	Difference
CNN	3.96/5	3.41/5	+0.56
XSum	3.84/5	2.58/5	+1.26

Visualizations

All visualizations are high-resolution PNG files saved in the results/ directory:

• Main Analysis Dashboard - 4-panel comparison showing all metrics
• Bias Analysis - Magnitude of self-preference bias
• CNN Distributions - Score distributions for CNN dataset
• XSum Distributions - Score distributions for XSum dataset

Data Files

• Summary Statistics CSV - All metrics in table format
• Detailed Analysis JSON - Complete statistical analysis
• Raw experiment results in results/cnn/ and results/xsum/

---

🔍 Key Conclusions

GPT-4.1-Nano demonstrates strong self-recognition capabilities:

1. ✅ Can detect own outputs with >92% accuracy
2. ✅ Prefers own outputs with >98% consistency
3. ✅ Recognizes own authorship with 25-31% higher probability
4. ✅ Rates own quality higher by 0.6-1.3 points

📚 Citation

@article{panickssery2024llm,
  title={LLM Evaluators Recognize and Favor Their Own Generations},
  author={Panickssery, Arjun and Bowman, Samuel R and Feng, Shi},
  journal={arXiv preprint arXiv:2404.13076},
  year={2024}
}