Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent

Xiaoyan Yu, Tongxu Luo, Yifan Wei, Fangyu Lei, Yiming Huang, Hao Peng, Liehuang Zhu

TL;DR

Neeko tackles Multi-Character Role-Playing (MCRP) by introducing dynamic, per-character LoRA blocks and a gating mechanism to seamlessly switch among multiple personas. It extends lifelong learning through fusion and expansion strategies to accommodate unseen characters while preserving prior knowledge. Comprehensive evaluation on Character-LLM-Data against ICL, RAG, and FT baselines demonstrates Neeko's superior character consistency, knowledge alignment, and dialogue stability, with competitive efficiency. The work advances MCRP methodology, proposes targeted evaluation metrics, and offers a practical framework for engaging, multi-character dialogue agents.

Abstract

Large Language Models (LLMs) have revolutionized open-domain dialogue agents but encounter challenges in multi-character role-playing (MCRP) scenarios. To address the issue, we present Neeko, an innovative framework designed for efficient multiple characters imitation. Unlike existing methods, Neeko employs a dynamic low-rank adapter (LoRA) strategy, enabling it to adapt seamlessly to diverse characters. Our framework breaks down the role-playing process into agent pre-training, multiple characters playing, and character incremental learning, effectively handling both seen and unseen roles. This dynamic approach, coupled with distinct LoRA blocks for each character, enhances Neeko's adaptability to unique attributes, personalities, and speaking patterns. As a result, Neeko demonstrates superior performance in MCRP over most existing methods, offering more engaging and versatile user interaction experiences. Code and data are available at https://github.com/weiyifan1023/Neeko.

Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent

TL;DR

Neeko tackles Multi-Character Role-Playing (MCRP) by introducing dynamic, per-character LoRA blocks and a gating mechanism to seamlessly switch among multiple personas. It extends lifelong learning through fusion and expansion strategies to accommodate unseen characters while preserving prior knowledge. Comprehensive evaluation on Character-LLM-Data against ICL, RAG, and FT baselines demonstrates Neeko's superior character consistency, knowledge alignment, and dialogue stability, with competitive efficiency. The work advances MCRP methodology, proposes targeted evaluation metrics, and offers a practical framework for engaging, multi-character dialogue agents.

Abstract

Large Language Models (LLMs) have revolutionized open-domain dialogue agents but encounter challenges in multi-character role-playing (MCRP) scenarios. To address the issue, we present Neeko, an innovative framework designed for efficient multiple characters imitation. Unlike existing methods, Neeko employs a dynamic low-rank adapter (LoRA) strategy, enabling it to adapt seamlessly to diverse characters. Our framework breaks down the role-playing process into agent pre-training, multiple characters playing, and character incremental learning, effectively handling both seen and unseen roles. This dynamic approach, coupled with distinct LoRA blocks for each character, enhances Neeko's adaptability to unique attributes, personalities, and speaking patterns. As a result, Neeko demonstrates superior performance in MCRP over most existing methods, offering more engaging and versatile user interaction experiences. Code and data are available at https://github.com/weiyifan1023/Neeko.
Paper Structure (48 sections, 8 equations, 4 figures, 16 tables)

This paper contains 48 sections, 8 equations, 4 figures, 16 tables.

Table of Contents

  1. Introduction
  2. Problem Scope
  3. Task Formulation: MCRP
  4. LoRA: Low-Rank Adapter
  5. Role-Playing with LoRA
  6. Methodology
  7. Role-Playing with Dynamic LoRA
  8. Role Selection with Gating Network
  9. Lifelong Role-Playing with LoRA Expansion
  10. Role-Incremental via Fusion
  11. Role-Incremental via Expansion
  12. Evaluation
  13. Character Consistency
  14. Character Behavior (CB).
  15. Character Utterance (CU).
  16. ...and 33 more sections

Figures (4)

  • Figure 1: The overall framework of Neeko. The Neeko framework encompasses three main phases: Agent Pre-Tuning, Role-Playing, and Incremental Learning. The Incremental Learning phase is achieved with two strategies: fusion and expansion.
  • Figure 2: Evaluation results across all metrics at the incremental stage. The horizontal comparisons among ICL (LLaMA-chat), RAG (LLaMA-chat), and FT (LoRA, Neeko) methods under the 7B parameter scale setting.
  • Figure 3: The distribution of three human evaluators on the responses generated by agents.
  • Figure 4: The interface of the program for human evaluation.