Research map / Large language models / Alignment and preferences
CAReDiO: Enhancing Cultural Alignment of LLM via Representativeness and Distinctiveness Guided Data Optimization
Jing Yao, Xiaoyuan Yi, Jindong Wang, Zhicheng Dou, Xing Xie
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 representativeness objective only vs raw data scoring (Only Rep) (ablation) | CAReDiO (Qwen2.5-7B-Instruct) · avg of CB-Easy, CB-Hard, Prism, GlobalOpinionQA | average: 58.05 -> 59.43 | Improved |
| add distinctiveness objective to representativeness-only (ablation) | CAReDiO (Qwen2.5-7B-Instruct) · avg of four benchmarks | average: 59.43 -> 61.83 | Improved |
| add representativeness objective to distinctiveness-only (ablation) | CAReDiO (Qwen2.5-7B-Instruct) · avg of four benchmarks | average: 58.50 -> 61.83 | Improved |
| replace role-play with CAReDiO data fine-tuning | Qwen2.5-7B-Instruct · Prism | rating (1-5): 3.364 -> 3.871 | Improved |
| fine-tune on CultureInstruction data | Gemma-3-27B-IT · CulturalBench-Hard | accuracy: 46.59 -> 18.19 | Hurt |
| fine-tune on CultureBank data | Qwen2.5-7B-Instruct · CulturalBench-Hard | accuracy: 38.90 -> 27.34 | Hurt |
| use CAReDiO instead of CultureLLM (WVS-augmented data) | Qwen2.5-7B-Instruct · GlobalOpinionQA / WVS | accuracy: 57.11 vs 56.23 (GlobalOpinionQA) | Hurt |
| fine-tune with CAReDiO data on GPT-4.1 | GPT-4.1 · CulturalBench-Hard | accuracy: 59.45 -> 63.54 | Improved |
| increase number of consensus individuals M (ablation) | CAReDiO · average over 15 cultures | performance | Improved |
| increase optimization rounds (ablation) | CAReDiO · average over 15 cultures | performance | Improved |
| increase training samples from 100 to 2,000 (ablation) | CAReDiO · average over 15 cultures | performance | Improved |
| aggregate multi-role consensus instead of single role (ablation) | GPT-5 · GlobalOpinionQA | accuracy: 58.63 -> 59.92 | Improved |
Closest papers by meaning
- CultureLLM: Incorporating Cultural Differences into Large Language Models (NeurIPS 2024 · similarity 0.92)
- CulturePark: Boosting Cross-cultural Understanding in Large Language Models (NeurIPS 2024 · similarity 0.90)
- BLEnD: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages (NeurIPS 2024 · similarity 0.88)
- Pangea: A Fully Open Multilingual Multimodal LLM for 39 Languages (ICLR 2025 · similarity 0.87)
- Attributing Culture-Conditioned Generations to Pretraining Corpora (ICLR 2025 · similarity 0.87)
- 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 0.86)
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.
