Identity-Driven Hierarchical Role-Playing Agents (2407.19412v1)

Abstract: Utilizing LLMs to achieve role-playing has gained great attention recently. The primary implementation methods include leveraging refined prompts and fine-tuning on role-specific datasets. However, these methods suffer from insufficient precision and limited flexibility respectively. To achieve a balance between flexibility and precision, we construct a Hierarchical Identity Role-Playing Framework (HIRPF) based on identity theory, constructing complex characters using multiple identity combinations. We develop an identity dialogue dataset for this framework and propose an evaluation benchmark including scale evaluation and open situation evaluation. Empirical results indicate the remarkable efficacy of our framework in modeling identity-level role simulation, and reveal its potential for application in social simulation.

• The paper introduces HIRPF, a framework that uses Lora and MoE to isolate and dynamically fuse identities for nuanced role simulation.
• Key evaluation using 20,685 dialogues shows enhanced identity fitting and fidelity compared to conventional prompt-based methods.
• The study demonstrates practical applications in social simulations by ensuring precise control and flexibility in portraying multiple identities.

Identity-Driven Hierarchical Role-Playing Agents: A Summary

The paper "Identity-Driven Hierarchical Role-Playing Agents" explores the integration of LLMs with identity theory to enhance role-playing capabilities. This approach addresses the limitations of conventional role-playing methods that rely on prompt-based strategies, which often lack precision, and fine-tuning methods, which struggle with flexibility.

Hierarchical Identity Role-Playing Framework (HIRPF)

The authors present the Hierarchical Identity Role-Playing Framework (HIRPF). This framework constructs identities from personality traits and professional categories, ensuring both identity isolation and explicit control. Leveraging techniques such as Lora and MoE, HIRPF isolates identities both intra-level and inter-level. The intra-level isolation utilizes specific Lora matrices for identities in the same category, while inter-level isolation distinguishes between identity categories by integrating them into model blocks alternately.

Explicit control is achieved through hard masking and soft routing. Hard masking ensures that only selected identities influence the model's output, allowing for flexible integration. This mechanism facilitates dynamic fusion of identities during role-play, supporting complex character simulations.

Identity Dialogue Dataset and Evaluation

The identity dialogue dataset developed for this project comprises 20,685 multi-turn dialogues. These dialogues cater to single and multiple identities, spanning personality traits and professions. The creation process involved generating dialogue content via ChatGPT, further enriched through re-annotation to ensure identity consistency.

Evaluation methods include a systematic benchmark with scale tests and open-ended situational assessments. The scale tests determine the fidelity of identity representation, while the situational tests measure how well the agents can integrate and express multiple identities.

Empirical Results and Implications

Empirical tests demonstrate the framework's ability to achieve more nuanced and realistic role simulation, particularly in identity-level role-playing. The results highlight superior identity fitting compared to models with equivalent parameters, especially in handling negative traits. Compared to prompt-based methods, HIRPF ensures sharper delineation and simulation of identities, proving beneficial in contexts like social simulations.

The framework's potential for practical applications extends to areas needing demographic or group-specific modeling. By focusing on identity-driven role-play, the research opens avenues for improved social governance simulations and other domains requiring nuanced human interaction emulations.

Future Developments

While demonstrating strengths in flexibility and precision, future enhancements could incorporate real-world conversational data to further refine the model's realism. Additionally, integrating retrieval-enhancement techniques may enable this framework to simulate more detailed and individualized character behaviors.

In conclusion, the Hierarchical Identity Role-Playing Framework sets a new standard in identity-driven role-play, offering a balanced approach to precision and adaptability. This framework enriches the capabilities of LLMs, paving the way for advanced social simulations and applications beyond conventional role-playing scopes.