Research Questions
- Can an LLM-based model learn and imitate an individual’s thoughts and judgments from their conversations?
- Can such individual-centered models be used in survey research?
- Is it feasible to simulate personalized opinions in a practical single-device environment (e.g., a 40GB GPU)?
Results
- The Doppelgänger model replicated individual opinions with high accuracy.
- Achieved 67% accuracy on a 5-point scale and 80% accuracy on a 3-point scale.
- Longer context windows (e.g., 3,000 tokens) and more training epochs increased accuracy.
- Outperformed commercial models such as GPT-3.5 Turbo and Gemini.
- Proven feasible on a single 40GB GPU.
Findings
-
The Doppelgänger model can reproduce opinions not only at the group level but also at the individual level with strong fidelity.
-
Surveyed LLMs (0.46 / 0.65 accuracy) and commercial models performed poorly at individual-level imitation.
-
Predictions based solely on metadata were weak and biased, underscoring the importance of conversational data.
-
The model learned effectively even with an average of 21 conversation samples per individual.
-
The approach provides a powerful tool for capturing individual-level heterogeneity and generating personalized responses.
-
LLM Models: 5
-
Synthetic Data: 5
-
Method: 5
-
Speed: 4
-
Ethics: 2
-
Accuracy: 5
-
Demographics: 2
If you would like to explore this research in more detail, click here to read the full paper.