Neeko is an innovative framework designed for efficient multi-character role-playing (MCRP) in large language models (LLMs). Unlike existing methods, Neeko employs a dynamic low-rank adapter (LoRA) strategy to seamlessly adapt to diverse characters. The framework is divided into three stages: agent pre-training, multiple characters playing, and character incremental learning. This approach effectively handles both seen and unseen roles, enhancing Neeko's adaptability to unique attributes, personalities, and speaking patterns. Neeko demonstrates superior performance in MCRP over most existing methods, offering more engaging and versatile user interaction experiences. The code and data are available at <https://github.com/weiyifan1023/Neeko>. - Formulate the novel task of multi-character role-playing (MCRP) agent learning and propose evaluation metrics tailored for this task. - Present Neeko, an incremental role-playing agent capable of playing multiple characters in long conversations and handling both seen and unseen characters. - Conduct extensive experiments using the Character-LLM-Data dataset and current LLMs like GPT-3.5 and LLaMA-2, demonstrating the challenging nature of MCRP and Neeko's superior performance. - **Task Formulation**: MCRP involves injecting the style of multiple characters into a language model to enhance conversation personalization. - **Low-Rank Adapter (LoRA)**: A parameter-efficient fine-tuning method that enables the adaptation of LLMs through lightweight modules. - **Pre-training**: Train non-overlapping LoRA blocks for each predefined character. - **Role-Playing**: Use a gating network to select and activate specific LoRA blocks for role-based instruction during inference. - **Incremental Learning**: Employ fusion and expansion strategies to handle new characters. - **Character Consistency**: Evaluate character behavior and utterance consistency. - **Knowledge Consistency**: Assess virtual and real knowledge, including hallucinatory knowledge. - **Dialogue Consistency**: Evaluate transfer, relevance, and stability in multi-turn dialogues. - **Pre-training Results**: Neeko outperforms baselines in both single-turn and multi-turn dialogues. - **Incremental Results**: Neeko achieves the best and second-best average performance with fusion and expansion strategies, respectively. - **Transfer Results**: Neeko outperforms all baseline methods in multi-role transfer. - **Consumption Results**: Neeko has lower memory usage and training time compared to other methods. - **Related Work**: Discusses previous efforts in role-playing agents and their limitations. - **Conclusion**: Highlights the superiority of Neeko in MCRP and its potential to advance the field. - **Limitations**: The MoE-like gate mechanism may result in less precise role representations, affecting overall performance.Neeko is an innovative framework designed for efficient multi-character role-playing (MCRP) in large language models (LLMs). Unlike existing methods, Neeko employs a dynamic low-rank adapter (LoRA) strategy to seamlessly adapt to diverse characters. The framework is divided into three stages: agent pre-training, multiple characters playing, and character incremental learning. This approach effectively handles both seen and unseen roles, enhancing Neeko's adaptability to unique attributes, personalities, and speaking patterns. Neeko demonstrates superior performance in MCRP over most existing methods, offering more engaging and versatile user interaction experiences. The code and data are available at <https://github.com/weiyifan1023/Neeko>. - Formulate the novel task of multi-character role-playing (MCRP) agent learning and propose evaluation metrics tailored for this task. - Present Neeko, an incremental role-playing agent capable of playing multiple characters in long conversations and handling both seen and unseen characters. - Conduct extensive experiments using the Character-LLM-Data dataset and current LLMs like GPT-3.5 and LLaMA-2, demonstrating the challenging nature of MCRP and Neeko's superior performance. - **Task Formulation**: MCRP involves injecting the style of multiple characters into a language model to enhance conversation personalization. - **Low-Rank Adapter (LoRA)**: A parameter-efficient fine-tuning method that enables the adaptation of LLMs through lightweight modules. - **Pre-training**: Train non-overlapping LoRA blocks for each predefined character. - **Role-Playing**: Use a gating network to select and activate specific LoRA blocks for role-based instruction during inference. - **Incremental Learning**: Employ fusion and expansion strategies to handle new characters. - **Character Consistency**: Evaluate character behavior and utterance consistency. - **Knowledge Consistency**: Assess virtual and real knowledge, including hallucinatory knowledge. - **Dialogue Consistency**: Evaluate transfer, relevance, and stability in multi-turn dialogues. - **Pre-training Results**: Neeko outperforms baselines in both single-turn and multi-turn dialogues. - **Incremental Results**: Neeko achieves the best and second-best average performance with fusion and expansion strategies, respectively. - **Transfer Results**: Neeko outperforms all baseline methods in multi-role transfer. - **Consumption Results**: Neeko has lower memory usage and training time compared to other methods. - **Related Work**: Discusses previous efforts in role-playing agents and their limitations. - **Conclusion**: Highlights the superiority of Neeko in MCRP and its potential to advance the field. - **Limitations**: The MoE-like gate mechanism may result in less precise role representations, affecting overall performance.