Kuairp 系列角色扮演模型技术报告

Model Features

Role-play tailored to game scenarios

A small dedicated model embeds role-play ability and domain world knowledge while preserving general tool-calling, to support diverse in-game play.

01

Small & Efficient

Goal 1 desc.

02

Prompt-Engineering Friendly

Goal 2 desc.

03

Built-in World Knowledge

Goal 3 desc.

04

General Tool-Calling Ability

Goal 4 desc.

Core Innovations

Three Core Innovations

From the SFT data pipeline and reinforcement learning to two-stage OPD on-policy distillation — with CDD at its core.

Instruction-injected user behavior

Feat 1 desc.

Reverse Profile Filtering (RPF)

Feat 2 desc.

Turn-count stratified sampling

Feat 3 desc.

PHASE 1 Domain Expert Training

General Model SFT RL Domain Expert

PHASE 2 OPD Alignment & Fusion

General Model Domain Expert

OPD-stage1 OPD-stage2 Expert with
General Ability

OPD Stage 1: inject format & style into the general model

Stage 1 desc.

OPD Stage 2: inject world knowledge into the general model

Stage 2 desc.

Dive into CDD

Cumulative-Divergence Decay (CDD)

The mechanism

Student

T1 T2 T3 T4 T5 T6

Teacher

d1

d2

d3

d4

d5

d6

c1

c2

c3

c4

c5

c6

w1

w2

w3

w4

w5

w6

Σ

l1

l2

l3

l4

l5

l6

di: the student-teacher divergence for token i
ci: the student-teacher divergence for token i
wi: the loss weight for token i
li: the weighted loss for token i
All li are summed to update the Student.

Evaluation Results

Domain ability peaks, general ability preserved

qwen3-8b (baseline) Kuairp 1.0^† M2-HER

85.14

92.08

89.22

60.00

72.50

60.00

61.48

97.44

98.89

97.57

99.12

99.62

93.75

100.00

96.40

1.90

32.86

n/a

24.83

24.78

n/a

Char-Consist.

Memory

Diversity

LangQuality

Length

World Knowledge^‡

Tool Calling^*‡

BFCL footnote.

Kuairp footnote.

M2-HER footnote.

More comparison data in our report

Domain role-play ability greatly improved

Point 1 desc.

Domain world knowledge internalized

Point 2 desc.

General tool-calling ability preserved

Point 3 desc.