English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

KuaiRP Series Role-playing Models Technical Report

Forum topic · 小凯 · 2026-09-13

Summary

The KuaiRP technical report (arXiv 2609.11127) presents the complete training solution behind Kuaishou's dedicated role-playing model series. The models target four objectives: simplified prompt engineering, highly stable output quality, built-in domain world knowledge, and efficient deployment at small parameter sizes. Injecting deep domain knowledge typically causes catastrophic forgetting of general agent capabilities, so the authors propose a multi-stage pipeline: a standardized character template plus an SFT data pipeline built from user behavior simulation and reverse profile filtering; an RL phase with a rule-based composite reward function that suppresses degradation such as length expansion and repetitive generation; and a novel self-distillation paradigm using Two-stage On-Policy Distillation (OPD) with Cumulative-Divergence Decay (CDD), where the domain-adapted model teaches the original base model. Experiments show KuaiRP models match state-of-the-art proprietary models in role-playing fidelity within target domains while recovering general agent capabilities at very low deployment cost.

KuaiRP Series Role-playing Models Technical Report

Paper: arXiv:2609.11127 Categories: cs.AI, cs.CL Authors: Yipeng Wang, Ziwei Zhang, Jiahui Zhang, Qi Gan, Kai Sheng Published: 2026-09-13

Overview

This paper introduces the complete technical solution for the KuaiRP series of role-playing models. The authors aim to achieve four core objectives for a dedicated role-playing model:

  • Simplified prompt engineering
  • Highly stable output quality
  • Built-in domain world knowledge
  • High-efficiency deployment with a small parameter size

    • The Core Challenge

    Effectively injecting deep domain knowledge often leads to severe catastrophic forgetting of the model's general agent capabilities. To overcome this trade-off, the authors propose a multi-stage training pipeline.

    Training Pipeline

    1. SFT stage: A standardized character template is designed, and an SFT data pipeline is constructed based on user behavior simulation and reverse profile filtering. 2. Reinforcement Learning stage: A rule-based composite reward function eliminates common degradation phenomena such as length expansion and repetitive generation. 3. Self-distillation stage: To recover general capabilities compromised during SFT and RL, the authors propose a novel self-distillation paradigm using Two-stage On-Policy Distillation (OPD) equipped with Cumulative-Divergence Decay (CDD). The domain-adapted model serves as the teacher and the original base model as the student, balancing deep domain knowledge injection with the preservation of general agent capabilities.

    Results

    Experimental results demonstrate that the KuaiRP models:

  • Match current state-of-the-art proprietary models in role-playing fidelity within the target domains
  • Successfully recover general agent capabilities
  • Maintain extremely low deployment costs

Tags

#role-playing-models#llm-training#reinforcement-learning#knowledge-distillation#catastrophic-forgetting#sft#arxiv#on-policy-distillation

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634798