Training LLMs to Predict National Survey Results
We asked GPT-4.1 to reason and predict the distribution of responses to a national survey. Then, we picked the best reasoning traces and fine-tuned Qwen-2.5-7B on those examples. The results show significant improvement - our fine-tuned model achieves an average Jensen-Shannon Divergence (a measure of error) of 0.1, compared to 0.2 for the base Qwen model. Importantly, this site only shows examples where both models formatted their answers properly, but our fine-tuned model only had proper formatting about 40% of the time. We are considering reinforcement learning as a way to teach formatting (and better reasoning). Please email any questions to awesel [at] stanford.edu! Super happy to talk about this project. You can read our technical write-up here.
Try clicking "New Random Example" to see how the models compare on different survey questions!
Selected References
- Costello, C., Guo, S., Goldie, A., & Mirhoseini, A. (2025, April 25). Think, Prune, Train, Improve: Scaling Reasoning without Scaling Models. arXiv:2504.18116. https://arxiv.org/abs/2504.18116
- DeepSeek-AI. (2025). DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv. https://arxiv.org/abs/2501.12948
- Durmus, E., et al. (2024). Towards measuring the representation of subjective global opinions in language models. arXiv. https://arxiv.org/abs/2306.16388
- Qwen Team. (2024, September). Qwen2.5: A Party of Foundation Models. Retrieved from https://qwenlm.github.io/blog/qwen2.5/
- Suh, J., Jahanparast, E., Moon, S., Kang, M., & Chang, S. (2025, February 24). Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions. arXiv:2502.16761