An Empirical Study on Enhancing Large Language Models for Long-Term Conversations in Korean

Large language models (LLMs) have shown strong performance in open-domain dialogue, yet they continue to struggle with long-term multi-session conversations (MSC), particularly in non-English languages such as Korean. In this work, we present a comprehensive empirical study on enhancing Korean MSC capabilities of LLMs through dataset construction, memory modeling, and parameter-efficient fine-tuning. We introduce an extended Korean MSC dataset that explicitly distinguishes between persona memory (long-term user attributes) and episode memory (short-term, event-driven information), enabling more effective memory management across sessions. Using this dataset, we evaluate LLM performance on three core MSC tasks: session summarization, memory update, and response generation. Our experiments reveal that Korean MSC is intrinsically more challenging than English MSC and that memory update and response generation require substantial reasoning ability. To address these challenges, we compare LoRA, DPO, MoE, CPT, Layer Tuning, and neuron-level tuning methods. Results consistently show that neuron tuning, guided by a novel language-specific neuron identification method based on activation scores and entropy, achieves superior performance and robustness, particularly in continual learning settings. Overall, our findings highlight neuron-level adaptation as an effective and interpretable approach for improving long-term conversational ability in low-resource languages.