Research map / Large language models / Alignment and preferences
On the Alignment of Large Language Models with Global Human Opinion
Yang Liu, Masahiro Kaneko, Chenhui Chu
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 |
|---|---|---|---|
| change language of prompt from English to the target language (no other steering) (ablation) | LLMs · Country opinion alignment, China and Germany | alignment score: China 0.8124 -> 0.8365; Germany 0.8408 -> 0.8608 (8-LLM avg) | Improved |
| add language steering to persona steering (ablation) | LLMs · China and Germany | alignment score: China 0.8357 -> 0.8593; Germany 0.8607 -> 0.8771 | Improved |
| add language steering to few-shot steering (ablation) | LLMs · China and Germany | alignment score: China 0.8601 -> 0.8771; Germany 0.8576 -> 0.8903 | Improved |
| add few-shot steering over persona steering (ablation) | LLMs · China and Germany | alignment score: China 0.8357 -> 0.8601; Germany 0.8607 -> 0.8576 | Mixed |
| add persona steering to English prompt | LLMs · China and Germany | alignment score: China 0.8124 -> 0.8357; Germany 0.8408 -> 0.8607 | Improved |
| language steering applied to a weak model (ablation) | GPT-3.5-Turbo · China, no steering | alignment score: 0.7816 -> 0.7976 (no significance marker) | No clear effect |
| language steering on some models for Germany (ablation) | Llama3 and GPT-5 · Germany, no steering | alignment score: Llama3 0.8441 -> 0.8494; GPT-5 0.8826 -> 0.8801 | No clear effect |
| evaluate on earlier WVS wave rather than latest wave | GPT-5, DeepSeek-R1 · WVS waves 5-7 | alignment score | Hurt |
| shuffle option order, reduce few-shot examples or change few-shot distribution (ablation) | LLMs · Prompt sensitivity | Pearson r with default: r between 0.9660 and 0.9980 | No clear effect |
Closest papers by meaning
- The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models (NeurIPS 2024 · similarity 0.91)
- Questioning the Survey Responses of Large Language Models (NeurIPS 2024 · similarity 0.89)
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.89)
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.89)
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.88)
- Group Preference Optimization: Few-Shot Alignment of Large Language Models (ICLR 2024 · similarity 0.88)
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.
