Research map / Large language models / Language, safety and interpretability
Linear Representations of Political Perspective Emerge in Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Probes on attention heads of three 7B chat LLMs linearly predict lawmakers' DW-NOMINATE scores, transfer to news outlet slant, and steer generated text's political stance via activation addition.
Problem
Whether LLMs hold linear representations of political perspective (liberal-conservative ideology) in activation space that can be used to monitor and steer generated text's ideological stance.
Approach
Fit ridge probes on individual attention head outputs to predict DW-NOMINATE scores of U.S. lawmakers from prompts simulating them; reuse these probes to score news-outlet slant and to steer outputs by adding scaled probe directions to head activations during generation.
Key results
- Top attention-head probes reach cross-validated Spearman ~0.85-0.86 on DW-NOMINATE; ensemble of K=32 heads reaches 0.87 (Llama-2), 0.864 (Mistral), 0.885 (Vicuna).
- Probes trained only on lawmakers transfer to news outlet Ad Fontes scores with Spearman 0.798 (Llama-2), 0.764 (Mistral), 0.720 (Vicuna).
- Steering with alpha from -30 to 30 shifts GPT-4o-rated slant monotonically; correlation with alpha reaches 0.607 for Llama-2-7b-chat at K in {64,80,96}, vs 0.396 Mistral and 0.381 Vicuna.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace linear probes with one-layer ReLU MLP probes (ablation) | Political-ideology probes on Llama-2-7B-chat · 116th U.S. Congress (held-out lawmakers) | Spearman correlation: 0.854 -> 0.855 | No clear effect |
| Replace linear probes with one-layer ReLU MLP probes (ablation) | Political-ideology probes on Vicuna-7B · 116th U.S. Congress (held-out lawmakers) | Spearman correlation: 0.861 -> 0.872 | Helped |
| Replace linear probes with one-layer ReLU MLP probes (ablation) | Political-ideology probes on Mistral-7B-Instruct · 116th U.S. Congress (held-out lawmakers) | Spearman correlation: 0.846 -> 0.838 | Hurt |
| Add linear interventions at early-to-middle layers (Layers 1–21), rather than middle-to-last layers (ablation) | Llama-2-7B-chat · Political-stance steering | Correlation between α and political slant: 0.540 | Helped |
| Add linear interventions at middle-to-last layers (Layers 22–32), rather than early-to-middle layers (ablation) | Llama-2-7B-chat · Political-stance steering | Effect on ideological expression | No clear effect |
| Increase the number of intervened-on attention heads (ablation) | Linear political-stance steering · Nine U.S. policy issues | Steering effectiveness | Helped |
| Increase intervention α (ablation) | Linear political-stance steering · Nine U.S. policy issues | Political slant: 0.607; 0.396; 0.381 | Helped |
| Increase intervention α (ablation) | Llama-2-7B-chat · ADVANCE Act | Political slant: ρ=0.479 | Helped |
| Increase intervention α (ablation) | Llama-2-7B-chat · 2023 UAW Strike | Political slant: ρ=0.470 | Helped |
Surprising findings
- Non-linear MLP probes do not substantially outperform linear probes, supporting linear representation of ideology.
- Interventions on middle-to-last layers have almost no steering effect, while early-to-middle layer interventions are effective.
Limitations
- Findings are based on relatively small 7B models and may not generalize to larger models.
- U.S.-centric; may not apply to less polarized political environments.
- Relies on GPT-4o as slant evaluator, with possible LLM-judge bias despite human validation on a small sample.
Experimental setup
Datasets: DW-NOMINATE scores (116th Congress) Ad Fontes Media outlet scores
Benchmarks: DW-NOMINATE lawmaker prediction Ad Fontes news outlet slant steering essays on nine policy issues
Metrics: Spearman correlation R^2 intraclass correlation mean slant rating
Baselines: one-layer MLP probe random permutation probe cubic-transformed label probe
Builds on: inference time intervention DW-NOMINATE linear probing
Closest papers by meaning
- Examining Alignment of Large Language Models through Representative Heuristics: the case of political stereotypes (ICLR 2025 · similarity 0.89)
- Quantifying Political Polarization through the Lens of Machine Translation and Vicarious Offense (AAAI 2024 · similarity 0.87)
- Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs (NeurIPS 2026 · similarity 0.86)
- When Neutral Summaries Are Not That Neutral: Quantifying Political Neutrality in LLM-Generated News Summaries (Student Abstract) (AAAI 2025 · similarity 0.85)
- The Power of LLM-Generated Synthetic Data for Stance Detection in Online Political Discussions (ICLR 2025 · similarity 0.84)
- Sparse Autoencoders for Hypothesis Generation (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.