Understanding and Mitigating Political Stance Cross-topic Generalization in Large Language Models

TL;DR

This paper investigates the internal neuron mechanisms behind political stance generalization in large language models and introduces a method to mitigate unintended cross-topic political influence.

Contribution

It identifies distinct political neuron types and proposes InhibitFT, a novel fine-tuning approach that reduces cross-topic stance generalization while maintaining performance.

Findings

Political neurons are consistent across models and datasets.

InhibitFT reduces cross-topic stance generalization by 20%.

Inhibiting 5% of neurons effectively mitigates the issue.

Abstract

Fine-tuning Large Language Models on a political topic will significantly manipulate their political stance on various issues and unintentionally affect their stance on unrelated topics. While previous studies have proposed this issue, there is still a lack of understanding regarding the internal representations of these stances and the mechanisms that lead to unintended cross-topic generalization. In this paper, we systematically explore the internal mechanisms underlying this phenomenon from a neuron-level perspective and how to mitigate the cross-topic generalization of political fine-tuning. Firstly, we propose Political Neuron Localization through Activation Contrasting (PNLAC) to identify two distinct types of political neurons: general political neurons, which govern stance across multiple political topics, and topic-specific neurons} that affect the model's political stance on individual topics. We find the existence of these political neuron types across four models and datasets through activation patching experiments. Leveraging these insights, we introduce InhibitFT, an inhibition-based fine-tuning method, effectively mitigating the cross-topic stance generalization. Experimental results demonstrate the robustness of identified neuron types across various models and datasets, and show that InhibitFT significantly reduces the cross-topic stance generalization by 20% on average, while preserving topic-specific performance. Moreover, we demonstrate that selectively inhibiting only 5% of neurons is sufficient to effectively mitigate the cross-topic stance generalization.