What question did this study set out to answer?

This research analyzes how large language models encode political ideology and its impact on decision-making across various tasks and modalities.

June 3, 2026Open Access

Probing Political Ideology in Large Language Models: How Latent Political Representations Generalize Across Task and Modality

Key Points

This research analyzes how large language models encode political ideology and its impact on decision-making across various tasks and modalities.
Identified political discourse dimensions using DW-NOMINATE scores through linear probing.
Conducted interventions on attention head activations to assess causal effects on ideological tasks.
Extended analysis to Vision-Language Models, mapping visual and textual political representations.
Demonstrated robust generalization of ideological effects in perception-based tasks, influencing confirmation bias.
Found notable asymmetries in simulated voting behaviors due to post-training guardrails, altering ideological spaces.
Illustrated successful cross-modal algorithmic steering affecting narrative and aesthetic in images without visual prompts.

Abstract

Large language models (LLMs) encode rich internal representations of political ideology, derived from the distributions of human internet data. Within the field of Computational Social Science, these models are increasingly utilized as "silicon samples" to simulate human populations, perform polling, and moderate content. However, it remains critically underexplored how these internal representations contribute to model decision-making, and how these latent semantic dimensions generalize across behavioral tasks and modalities. This thesis systematically investigates the causal role of learned ideological directions via inference-time interventions on attention head activations. First, I identify a liberal-conservative discourse dimension via linear probing using DW-NOMINATE scores and demonstrate its causal effects across text tasks, including political bias detection, simulated voting preferences, and ideological text neutralization. My findings reveal robust generalization in perception-based tasks, allowing algorithmic manipulation of the model's confirmation bias. However, I document notable asymmetries in simulated voting behaviors, exposing how post-training alignment guardrails (RLHF) can alter ideological representation spaces. Second, I extend this investigation into the multimodal frontier, illustrating that modern Vision-Language Models (VLMs) map visual political signifiers natively onto corresponding liberal-conservative textual axes. I execute zero-shot image token probing on Qwen3-VL utilizing the DW-NOMINATE text direction. Finally, I demonstrate cross-modal algorithmic steering on the generic LLaMA textual backbone of Janus-Pro, causally altering the narrative, demographic, and aesthetic generation of images without any explicit visual prompt engineering. This work highlights the notable risks of using ideologically biased foundation models for politically sensitive sociological tasks and underscores the necessity of managing latent political dimensions in unified multimodal systems.

Read Full Paperexternally

Ask AI

Helpful

Bookmark

View Full Paper