KuaiRP Series Role-Playing Models Technical Report

Summary

This technical report presents the complete solution for the KuaiRP series, a set of dedicated role-playing models designed to simplify prompt engineering, stabilize output quality, incorporate domain world knowledge, and support efficient deployment with relatively few parameters. The authors identify a central trade-off: injecting deep domain knowledge can cause catastrophic forgetting of general agent capabilities. Their training pipeline first uses a standardized character template and supervised fine-tuning data built through simulated user behavior and reverse profile filtering. During reinforcement learning, a rule-based composite reward function is used to reduce length expansion, repetitive generation, and other common quality degradations. The authors then introduce Two-stage On-Policy Distillation with Cumulative-Divergence Decay to restore general capabilities affected by supervised fine-tuning and reinforcement learning. In this process, the domain-adapted model serves as teacher and the original base model as student. Experiments reported in the paper show that KuaiRP matches current proprietary state-of-the-art models in role-playing fidelity within the target domains, while recovering general agent capabilities and maintaining very low deployment costs.