Multi-dimensional Evaluation of Character-Authentic Dialogue Models Learned from Question-Answer Data

Abstract

Character-authentic dialogue remains challenging for large language models (LLMs) due to limited character-specific data, generic-style collapse, and hallucinations regarding persona facts. Our work presents a comparative evaluation of several learning strategies for character dialogue grounded in question–answer (QA) data, comparing zero/few-shot prompting, supervised fine-tuning (SFT), direct preference optimization (DPO), and a hybrid approach that integrates retrieval-augmented character profiles and knowledge with policy optimization. Using both single-turn and multi-turn settings, we assess multiple dimensions central to character dialogue quality: reproducibility, diversity, hallucination, and character authenticity. Results show that SFT excels in reproducibility and hallucination reduction but tends to shorten and simplify outputs, thereby reducing diversity and authenticity. DPO improves stylistic fidelity and authenticity but depends strongly on externalized character knowledge to limit hallucinations. The hybrid variant that combines character-knowledge retrieval with DPO achieves the best overall balance, delivering strong authenticity while maintaining factual consistency and competitive reproducibility in both single- and multi-turn dialogues. We further analyze the model’s sensitivity to knowledge retrieval and response-length effects and discuss trade-offs among optimization targets that inform practical design choices for developing faithful and engaging character agents trained from scalable QA resources.

Resources

Details