This paper details the technical solution for the KuaiRP series of role-playing models. The core objectives are simplified prompt engineering, stable output quality, built-in domain world knowledge, and high-efficiency deployment with a small parameter size. The approach addresses the common trade-off between domain adaptation and general agent capabilities. A multi-stage training pipeline is proposed, including a standardized character template and an SFT data pipeline based on user behavior simulation and reverse profile filtering.
Reinforcement Learning (RL) is then utilized with a rule-based composite reward function to mitigate degradation phenomena such as length expansion and repetitive generation. A novel self-distillation paradigm, Two-stage On-Policy Distillation (OPD) equipped with Cumulative-Divergence Decay (CDD), is employed to recover general agent capabilities. This utilizes a domain-adapted model as the teacher and the original base model as the student.
Experimental results show that the KuaiRP models achieve role-playing fidelity comparable to proprietary models within specified domains. Furthermore, the models maintain extremely low deployment costs while preserving general agent capabilities. This approach balances deep domain knowledge injection with the preservation of general agent capabilities.
Source: https://arxiv.org/abs/2609.11127