Linear Representations of Political Perspective Emerge in Large Language Models

TL;DR

This research suggests that LLMs possess a high-level linear representation of American political ideology and that by leveraging recent advances in mechanistic interpretability, they can identify, monitor, and steer the subjective perspective underlying generated text.

Abstract

Large language models (LLMs) have demonstrated the ability to generate text that realistically reflects a range of different subjective human perspectives. This paper studies how LLMs are seemingly able to reflect more liberal versus more conservative viewpoints among other political perspectives in American politics. We show that LLMs possess linear representations of political perspectives within activation space, wherein more similar perspectives are represented closer together. To do so, we probe the attention heads across the layers of three open transformer-based LLMs (Llama-2-7b-chat, Mistral-7b-instruct, Vicuna-7b). We first prompt models to generate text from the perspectives of different U.S. lawmakers. We then identify sets of attention heads whose activations linearly predict those lawmakers' DW-NOMINATE scores, a widely-used and validated measure of political ideology. We find that highly predictive heads are primarily located in the middle layers, often speculated to encode high-level concepts and tasks. Using probes only trained to predict lawmakers' ideology, we then show that the same probes can predict measures of news outlets' slant from the activations of models prompted to simulate text from those news outlets. These linear probes allow us to visualize, interpret, and monitor ideological stances implicitly adopted by an LLM as it generates open-ended responses. Finally, we demonstrate that by applying linear interventions to these attention heads, we can steer the model outputs toward a more liberal or conservative stance. Overall, our research suggests that LLMs possess a high-level linear representation of American political ideology and that by leveraging recent advances in mechanistic interpretability, we can identify, monitor, and steer the subjective perspective underlying generated text.

Figures (14)

• Figure 1: Excerpts from essays generated by Mistral-7b-instruct on policy issues (e.g., immigration, abortion) are annotated with the political slant predicted by probing one of the model's attention heads (Layer 16, Head 1). This attention head was among the most predictive heads with the highest Spearman correlation in predicting the political ideology of U.S. lawmakers. Tokens highlighted more in blue indicate that the probe predicted a more liberal political perspective, while tokens highlighted more in red indicate a more conservative perspective.
• Figure 2: Predictive performance of linear probes for all attention heads across all layers in Llama-2-7b-chat, Mistral-7b-instruct, and Vicuna-7b. Each row (i.e., $y$-axis) represents each layer of the model from the bottom (layers close to the input layer) to the top (layers close to the output layer). Each column (i.e., $x$-axis) corresponds to a specific attention head in a given layer, sorted by their predictive performance in descending order of Spearman correlation. Darker versus lighter shades indicate higher versus lower Spearman correlation, meaning the attention head was more or less predictive of lawmakers' political ideology (i.e., DW-NOMINATE scores).
• Figure 3: Ideological perspectives of U.S. lawmakers and news media as predicted by the activations of the $K=32$ most predictive attention heads of Llama-2-7b-chat. Negative values correspond to left-leaning perspectives, while positive values correspond to right-leaning perspectives. The $x$-axis represents the predicted political slant ($\widehat{y}_{K=32}^{\mathsmaller{(i)}}$) for each entity (i.e., lawmakers or news media). The $y$-axis represents the previously validated ideological scores (DW-NOMINATE or Ad Fontes Media scores). See \ref{['fig:ideo_all_models', 'fig:ideo_all_models_news_media']} for the complete results across all models.
• Figure 4: Trained probes can be used effectively to steer the political slant of generated text; see (a). Steering is more reliable for certain policy issues, but has a positive effect for all; see (b). LLMs steered toward more liberal positions on certain policy issues tend to produce longer essays; see (c).
• Figure A1: Ideological perspectives of U.S. lawmakers as captured by the activation space in Llama-2-7b-chat, Mistral-7b-instruct, and Vicuna-7b. Negative values correspond to left-leaning perspectives, while positive values correspond to right-leaning perspectives. Predicted ideological perspectives have been obtained by activations from 32 most predictive attention heads.