Real LLaMA-3.1-8B-Instruct's political orientation is causally steerable via per-attention-head linear center-of-mass (difference-of-means) directions that shift Political Compass responses toward alternative ideological positions, in English and, to a lesser degree, in four other tested languages

- Gurgurov, Trinley, Vykopal, van Genabith, Ostermann & Zamparelli evaluate political orientation in seven real instruction-tuned open-source LLMs (LLaMA-3.1-8B/70B, Qwen-3-8B/14B/32B, Aya-Expanse-8B/32B) across 14 languages using the Political Compass Test, finding larger models consistently shift toward libertarian-left positions. [gurgurov-etal-2025-multilingual-political-views] - Focusing on LLaMA-3.1-8B (an instruction-tuned model), they train binary linear probes on the output of every attention head (liberal vs. conservative, English-only training data; best probe accuracy 0.90 at layer 30, head 19) and build per-head steering vectors by the center-of-mass (difference-of-class-means, i.e. DiffMean) method; adding them to the top-K most responsive heads (K=512, alpha=20 in the main figure) shifts Political Compass responses toward the liberal or conservative target across all 11 paraphrases in English, and the English-derived vectors also shift outputs in Turkish, Romanian, Slovenian and French 'to a certain degree' (with more refusals/irrelevant answers in non-English languages). The steering is tested on this single model, with limited hyperparameter exploration, and the directions are per-attention-head rather than a single residual-stream direction. [gurgurov-etal-2025-multilingual-political-views]