Generalization to Political Beliefs from Fine-Tuning on Sports Team Preferences
Owen Terry
TL;DR
The paper investigates cross-domain generalization when fine-tuning large language models on a simple, narrow-domain task—sports-team preferences—by training coastal and southern variants of Qwen2.5-32B-Instruct and comparing them to a base model on nine political belief questions rated on a 0-9 scale. It analyzes top-token distributions and collects elaborations for radical answers, finding broader distributions and some incoherent or contradictory explanations in the fine-tuned models, but no consistent liberal or conservative tilt relative to the base model. The results suggest that learning from narrow, opinionated data can produce seemingly unrelated behavioral changes and highlight the need for deeper causal and interpretability studies, including the role of prompting and training configuration.
Abstract
Fine-tuned LLMs often exhibit unexpected behavior as a result of generalizing beyond the data they're shown. We present results in which an LLM fine-tuned to prefer either coastal sports teams or Southern sports teams adopt political beliefs that diverge significantly from those of the base model. While we hypothesized that the coastal model would become more liberal and the southern model would become more conservative, we find that their responses are usually similar to each other, without a clear-cut liberal or conservative bias. In addition to asking the models for numerical ratings of agreement with relevant political statements, we ask them to elaborate on their more radical answers, finding varying degrees of willingness to justify themselves. Further work is needed to understand the mechanisms by which fine-tuning on simple, narrow datasets leads to seemingly unrelated changes in model behavior.