Research map / Large language models / Alignment and preferences
CultureRL: Internalizing Cultural Principles in Large Language Models via Norm-Driven Reinforcement Learning
Weixiang Zhao, Haozhen Li, Yanyan Zhao, Haixiao Liu, Biye Li, Ting Liu, Bing Qin
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 |
|---|---|---|---|
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-IRQ · WVS (IRQ) | 1 - JSD: 82.84 -> 86.35 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-IRQ · VSM13 (IRQ) | Euclidean distance: 77.16 -> 62.24 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-IRQ · Content moderation (IRQ) | F1: 43.92 -> 49.34 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-TUR · WVS (TUR) | 1 - JSD: 78.50 -> 79.98 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-TUR · VSM13 (TUR) | Euclidean distance: 129.52 -> 111.25 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-TUR · Content moderation (TUR) | F1: 50.61 -> 57.37 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-USA · WVS (USA) | 1 - JSD: 76.63 -> 76.72 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-USA · VSM13 (USA) | Euclidean distance: 173.25 -> 131.40 | Improved |
| add norm-cluster-based reward (cluster-specific norm rules) instead of all norm rules without clustering (ablation) | CultureRL-USA · Content moderation (USA) | F1: 38.61 -> 40.98 | Improved |
| use norm-cluster-based reward instead of top-5 retrieved norm rules per sample (ablation) | CultureRL-IRQ · WVS (IRQ) | 1 - JSD: 72.34 -> 86.35 | Improved |
| use norm-cluster-based reward instead of top-5 retrieved norm rules per sample (ablation) | CultureRL-IRQ · VSM13 (IRQ) | Euclidean distance: 155.03 -> 62.24 | Improved |
| use norm-cluster-based reward instead of top-5 retrieved norm rules per sample (ablation) | CultureRL-IRQ · Content moderation (IRQ) | F1: 46.37 -> 49.34 | Improved |
Closest papers by meaning
- CultureLLM: Incorporating Cultural Differences into Large Language Models (NeurIPS 2024 · similarity 0.89)
- CulturePark: Boosting Cross-cultural Understanding in Large Language Models (NeurIPS 2024 · similarity 0.88)
- RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style (ICLR 2025 · similarity 0.88)
- Artificial Generational Intelligence: Cultural Accumulation in Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Meta-Learning Preferences for Multilingual LLM Alignment (NeurIPS 2026 · similarity 0.87)
- Training-Free Cultural Alignment of Large Language Models via Persona Disagreement (NeurIPS 2026 · similarity 0.87)
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.
