Enhancing Role-Playing Language Models via Similar-Scene Generation

View at publisher Summarise with VIRAT
Authors: Hong Yan
Cited by: 0
References: 18

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

With the rapid advancement of large language models, role-playing language agents (RPLAs) have attracted increasing attention for their ability to simulate diverse personas and enable personalized human-AI interaction. However, existing approaches to building RPLAs often require substantial human effort for manual data construction or annotation, and are constrained by the limited scale of available character-specific dialogue data, which hinders their scalability and practical deployment. To address these challenges, we propose a novel framework for constructing customized role-playing language agents based on LLMs. Our approach integrates two key components: (1) diverse instruction generation, achieved through an iterative self-expansion process combining seed QA examples with similarity-based filtering to ensure both reliability and diversity; (2) similar-scene-enhanced response generation, which retrieves semantically related QA samples as contextual guidance to produce consistent and persona-aligned responses. Comprehensive experiments on representative characters from the RoleBench dataset demonstrate that our method significantly improves performance across multiple evaluation dimensions, including personality consistency, background integration, language style, and overall response quality. Further ablation studies confirm that both the similar-scene enhancement and dual filtering mechanisms are crucial for achieving robust and natural role-playing behavior. Our work provides a practical and extensible paradigm for developing high-fidelity, multi-persona conversational agents, advancing the frontier of personalized dialogue systems and human-like language generation.

Fields of Study & Subjects

Speech and dialogue systems Topic Modeling AI in Service Interactions

Publication details

Journal / source
International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
Pages
1237–1243
Publication date
16 January 2026
Language
EN