Mind the Gap: Pitfalls in Aligning Large Language Models with Asian Public Opinion

The researchers conducted a multilingual cultural alignment audit covering India, East Asia, and Southeast Asia on five models—GPT-4o-Mini, Gemini-2.5-Flash, Llama 3.2, Mistral, and Gemma 3—using religion as an entry point for examining broader alignment. The team used log-probs/logits to analyze the internal representations of each model and compared the models' opinion distributions with real public attitudes, finding that these models largely align with public opinion on broad social issues but consistently fail to accurately represent religious views, especially those of minority groups, and often amplify negative stereotypes. Lightweight interventions such as demographic feature prompts and native-language prompts only partially mitigate and cannot eliminate these cultural gaps. Downstream evaluations on bias benchmarks such as CrowS-Pairs, IndiBias, ThaiCLI, and KoBBQ also show persistent harm and underrepresentation in sensitive scenarios, on which basis the authors emphasize the need for systematic regional audits to ensure fair global deployment of LLMs.