Diagnosing and Mitigating LLM Failures in Recognizing Culturally Specific Korean Names: An Error-Driven Prompting Framework

Abstract

As large language models (LLMs) improve in understanding and reasoning, they are increasingly used in privacy protection tasks such as de-identification, privacy-sensitive text generation, and entity obfuscation. However, these applications depend on an essential requirement: the accurate identification of personally identifiable information (PII). Compared with template-based PII that follows clear structural patterns, name-related PII depends much more on cultural and pragmatic context, which makes it harder for models to detect and raises higher privacy risks. Although recent studies begin to address this issue, existing work remains limited in language coverage, evaluation granularity, and the depth of error analysis. To address these gaps, this study proposes an error-driven framework that integrates diagnosis and intervention. Specifically, the framework introduces a method called Error-Driven Prompt (EDP), which transforms common failure patterns into executable prompting strategies. It further explores the integration of EDP with general advanced prompting techniques such as Chain-of-Thought (CoT), few-shot learning, and role-playing. In addition, the study constructed K-NameDiag, the first fine-grained evaluation benchmark for Korean name-related PII, which includes twelve culturally sensitive subtypes designed to examine model weaknesses in real-world contexts. The experimental results showed that EDP improved F1-scores in the range of 6 to 9 points across three widely used commercial LLMs, namely Claude Sonnet 4.5, GPT-5, and Gemini 2.5 Pro, while the Combined Enhanced Prompt (CEP), which integrates EDP with advanced prompting strategies, resulted in different shifts in precision and recall rather than consistent improvements. Further subtype-level analysis suggests that subtypes reliant on implicit cultural context remain resistant to correction, which shows the limitations of prompt engineering in addressing a model’s lack of internalized cultural knowledge.

Affiliated Institutions

• Yonsei University KR

Related Publications