Research map / Large language models / AI and society
Quantifying the Salience of Geo-Cultural Values for Pluralistic Safety Alignment
Arkadiy Saakyan, Charvi Rastogi, Lora Aroyo
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 cultural zone as fixed effect to base rater/item model | multilevel safety rating model · 6 of 8 safety datasets | model fit (Delta AIC): Delta AIC -6.23 to -195.93 | Improved |
| add demographics (ethnicity, age, gender) as fixed effects | multilevel safety rating model · 8 safety datasets | model fit (Delta AIC): Delta AIC -6.85 to -45.11 | Improved |
| add cultural zones on top of demographics (ablation) | demographics-only multilevel model · 6 of 8 datasets (not DIVE, NLPos) | Delta AIC: Delta AIC -3.97 to -179.88 | Improved |
| add culture x demographics interaction terms (ablation) | D+CZ multilevel model · 7 of 8 datasets | Delta AIC / LRT: D3 only: -94.91, p < 0.001 | No clear effect |
| use culturally sensitive item triage with fine-tuned LMs (safe vs sensitive instead of safe vs unsafe) | DeBERTa and Gemma-3-4B classifiers · D3 | F1: about 16% drop DeBERTa; about 14% drop Gemma | Hurt |
| scale from DeBERTa-Large (435M) to Gemma-3 4B decoder (ablation) | quadrant rating emulation · DICES-990, CREHate, D3 | F1 | No clear effect |
| prompt reasoning LLMs (Gemini-3 Flash, GPT-5 Nano) to emulate cultural quadrants | zero-shot LLM raters · D3 | F1 vs always-unsafe baseline | No clear effect |
| train on safe-vs-sensitive task and transfer to safe-vs-unsafe (ablation) | fine-tuned classifiers · D3 | F1 cross-task | Improved |
| train classifier on D3 and test on other datasets (ablation) | fine-tuned classifiers · CREHate, DICES-990 | F1 | No clear effect |
| raise Siq threshold from 0.5 to 0.7 (ablation) | cultural sensitivity score · 6 datasets mean | culturally sensitive item rate: 10.53 -> 3.43 | Hurt |
Closest papers by meaning
- SafeWorld: Geo-Diverse Safety Alignment (NeurIPS 2024 · similarity 0.91)
- Whose View of Safety? A Deep DIVE Dataset for Pluralistic Alignment of Text-to-Image Models (NeurIPS 2025 · similarity 0.88)
- ROK-FORTRESS: Measuring the Effect of Geopolitical Transcreation for National Security and Public Safety (NeurIPS 2026 · similarity 0.87)
- Culture Affordance Atlas: Reconciling Object Diversity Through Functional Mapping (AAAI 2026 · similarity 0.86)
- Large Language Models are Geographically Biased (ICML 2024 · similarity 0.85)
- Controllable Safety Alignment: Inference-Time Adaptation to Diverse Safety Requirements (ICLR 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.
