Research map / Large language models / Language, safety and interpretability
Fine-Grained Interpretation of Political Opinions in Large Language Models
Jingyu Hu, Mengyue Yang, Mengnan Du, Weiru Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use supervised linear probing instead of unsupervised RepE for detection | political concept vectors · in-distribution detection | detection accuracy: 0.8434 -> 0.9646 on Llama3-8B | Improved |
| use linear probing instead of CAA for detection | political concept vectors · in-distribution detection | detection accuracy: 0.9229 -> 0.9646 on Llama3-8B | Improved |
| use larger model within family (Llama3-3B to 8B) | linear probing · in-distribution detection | detection accuracy: 0.9452 -> 0.9646 | Improved |
| use fine-grained four-dimension probes instead of single-axis left-right probe | linear probing · OOD political statements, Llama3-8B | detection accuracy | Improved |
| go from early (layer 8) to late (layer 28) layers | RepE concept vectors · concept confound correlation | correlation between Left-Equality and Right-Authority: 0.85 -> -0.17 | Improved |
| add prompt template P2 (The leaning is) to base statements (ablation) | linear probing · OOD economic dimension | detection accuracy | Improved |
| intervene at middle layers instead of very early or late layers (ablation) | concept vector steering · Llama3-8B economic dataset | representation shift | Improved |
| raise steering strength alpha to 2.5 (ablation) | RepE right-leaning intervention · Mistral-7B inheritance question | response coherence | Hurt |
| add right-leaning intervention at layers 15 to 25 | RepE vector · Mistral-7B next-token intention | output opinion | Mixed |
Closest papers by meaning
- Linear Representations of Political Perspective Emerge in Large Language Models (ICLR 2025 · similarity 0.88)
- Examining Alignment of Large Language Models through Representative Heuristics: the case of political stereotypes (ICLR 2025 · similarity 0.86)
- Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses (NeurIPS 2026 · similarity 0.86)
- Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs (NeurIPS 2026 · similarity 0.85)
- When Neutral Summaries Are Not That Neutral: Quantifying Political Neutrality in LLM-Generated News Summaries (Student Abstract) (AAAI 2025 · similarity 0.84)
- The Power of LLM-Generated Synthetic Data for Stance Detection in Online Political Discussions (ICLR 2025 · similarity 0.83)
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.
