2608.04205v1 Aug 04, 2026 cs.AI

MatrAIx: 83억 개의 가상 에이전트를 활용한 세계 시뮬레이션

MatrAIx: Simulating the World with 8.3 Billion Persona Agents

Ziyun Zhang
Ziyun Zhang
Citations: 0
h-index: 0
Yunhan Zhao
Yunhan Zhao
Citations: 259
h-index: 6
Pengrui Han
Pengrui Han
Citations: 147
h-index: 8
Aravind Mohan
Aravind Mohan
Citations: 13
h-index: 3
Yijun Wang
Yijun Wang
Citations: 64
h-index: 4
Dawn Song
Dawn Song
Citations: 10
h-index: 2
Xuandong Zhao
Xuandong Zhao
UC Berkeley
Citations: 3,733
h-index: 30
Zonglin Di
Zonglin Di
Citations: 177
h-index: 5
Hanwen Xing
Hanwen Xing
Citations: 127
h-index: 3
Bowen Jiang
Bowen Jiang
Citations: 111
h-index: 5
Fang Wu
Fang Wu
Citations: 90
h-index: 4
Zidi Xiong
Zidi Xiong
Citations: 105
h-index: 4
Heming Liu
Heming Liu
Citations: 4
h-index: 1
Jian Hou
Jian Hou
Citations: 25
h-index: 3
Bowen Yang
Bowen Yang
University of Science and Technology of China
Citations: 1,032
h-index: 5
S. Schmidgall
S. Schmidgall
Citations: 588
h-index: 7
Wenhao Chai
Wenhao Chai
Citations: 420
h-index: 8
Fan Nie
Fan Nie
Citations: 81
h-index: 6
Zhiwei Zhang
Zhiwei Zhang
Citations: 4
h-index: 1
Shivam Patel
Shivam Patel
Citations: 6
h-index: 1
James Zou
James Zou
Citations: 54
h-index: 3
Abinitha Gourabathina
Abinitha Gourabathina
Citations: 19
h-index: 2
Qianfeng Wen
Qianfeng Wen
Citations: 45
h-index: 4
Yuanda Xu
Yuanda Xu
Citations: 80
h-index: 5
Asu Ozdaglar
Asu Ozdaglar
Citations: 61
h-index: 4
Yixuan He
Yixuan He
Citations: 58
h-index: 4
Tianyi Peng
Tianyi Peng
Citations: 19
h-index: 3
Muhammad Ahmed Mohsin
Muhammad Ahmed Mohsin
Citations: 4
h-index: 1
Emily Fox
Emily Fox
Citations: 42
h-index: 2
K. Hamidieh
K. Hamidieh
Citations: 286
h-index: 7
Huangyuan Su
Huangyuan Su
Citations: 31
h-index: 2
Mitchell Gordon
Mitchell Gordon
Citations: 307
h-index: 4
Bing Hu
Bing Hu
Citations: 17
h-index: 2
Yuexing Hao
Yuexing Hao
Citations: 17
h-index: 3
Jiayuan Ding
Jiayuan Ding
Citations: 292
h-index: 4
Xiaomin Li
Xiaomin Li
Citations: 129
h-index: 6
Zheyuan Deng
Zheyuan Deng
Citations: 4
h-index: 1
Jiawei Zhang
Jiawei Zhang
Citations: 622
h-index: 11
Y. Kim
Y. Kim
Citations: 646
h-index: 8
Chanwoo Park
Chanwoo Park
Citations: 308
h-index: 5
M. Zitnik
M. Zitnik
Citations: 26,045
h-index: 62
H. Jeong
H. Jeong
Citations: 440
h-index: 5
Sankalp Jajee
Sankalp Jajee
Citations: 1
h-index: 1
Philip H. S. Torr
Philip H. S. Torr
Citations: 75,314
h-index: 111
Yilun Du
Yilun Du
Citations: 9
h-index: 2
Dianzhuo Wang
Dianzhuo Wang
Citations: 34
h-index: 4
Yongye Su
Yongye Su
Citations: 18
h-index: 1
Brihi Joshi
Brihi Joshi
Citations: 34
h-index: 4
Jiuyao Lu
Jiuyao Lu
Citations: 22
h-index: 2
Jintao Huang
Jintao Huang
Citations: 55
h-index: 2
Yifan Liu
Yifan Liu
Citations: 0
h-index: 0
Xiaoyi Liu
Xiaoyi Liu
Citations: 0
h-index: 0
Yi-Mei Fan
Yi-Mei Fan
Citations: 0
h-index: 0
Ko-Hui Wu
Ko-Hui Wu
Citations: 1
h-index: 1
Ru-Ping Gao
Ru-Ping Gao
Citations: 0
h-index: 0
S. Kapoor
S. Kapoor
Citations: 0
h-index: 0
Isha Gupta
Isha Gupta
Citations: 0
h-index: 0
Wei-Wei Xiao
Wei-Wei Xiao
Citations: 0
h-index: 0
Run Zhang
Run Zhang
Citations: 0
h-index: 0
Mihir Kulshreshtha
Mihir Kulshreshtha
Citations: 0
h-index: 0
Yuxin Xiao
Yuxin Xiao
Citations: 55
h-index: 4
Zuxin Liu
Zuxin Liu
Citations: 951
h-index: 15
L. Chen
L. Chen
Citations: 2,723
h-index: 14
Ethan Ye
Ethan Ye
Citations: 614
h-index: 4
Jason Xie
Jason Xie
Citations: 626
h-index: 1
A.M. Richmond
A.M. Richmond
Citations: 0
h-index: 0
Wei Ding
Wei Ding
Citations: 0
h-index: 0
Emre Okçular
Emre Okçular
Citations: 0
h-index: 0
D. Mathew
D. Mathew
Citations: 0
h-index: 0
Rana Muhammad Shahroz Khan
Rana Muhammad Shahroz Khan
Vanderbilt University
Citations: 72
h-index: 5
Zhe Peng
Zhe Peng
Citations: 0
h-index: 0
Xinyang Han
Xinyang Han
Citations: 70
h-index: 3
Zhenting Qi
Zhenting Qi
Citations: 9
h-index: 2
Xue-Wei Pan
Xue-Wei Pan
Citations: 0
h-index: 0
H. Ramesh
H. Ramesh
Citations: 4
h-index: 1
K. Alhamoud
K. Alhamoud
Citations: 535
h-index: 10
Yepeng Huang
Yepeng Huang
Citations: 486
h-index: 8
Yonghe Wang
Yonghe Wang
Citations: 0
h-index: 0
Alex Gu
Alex Gu
Citations: 6
h-index: 1
Yuchuan Wang
Yuchuan Wang
Citations: 0
h-index: 0
Akshay Paruchuri
Akshay Paruchuri
Stanford University
Citations: 292
h-index: 8
Brenna Li
Brenna Li
Citations: 151
h-index: 7
Hejie Cui
Hejie Cui
Citations: 160
h-index: 4
Chaosheng Dong
Chaosheng Dong
Citations: 20
h-index: 3
P. Bhattacharya
P. Bhattacharya
Citations: 948
h-index: 11
Paul Pu Liang
Paul Pu Liang
Citations: 109
h-index: 4
Prasanna B. Tambe
Prasanna B. Tambe
Citations: 3,289
h-index: 20
Shirley Huang
Shirley Huang
Citations: 0
h-index: 0
Jing Tang
Jing Tang
Citations: 0
h-index: 0
Yucheng Lu
Yucheng Lu
Citations: 19
h-index: 2
Qian Zhu
Qian Zhu
Citations: 0
h-index: 0
Ziheng Wang
Ziheng Wang
Citations: 0
h-index: 0
Jiahao Wang
Jiahao Wang
Citations: 0
h-index: 0

인공지능 시스템 및 디지털 제품에 대한 인간 평가에는 비용이 많이 들고, 속도가 느리며 확장성이 떨어집니다. 오프라인 평가는 더 높은 확장성을 제공하지만, 종종 인간의 다양성과 상호 작용 방식을 간과합니다. 따라서 우리는 다양한 사용자를 대상으로 인공지능 시스템 및 디지털 제품을 테스트하기 위한 대규모 시뮬레이션 사용자 평가 인프라인 MatrAIx를 소개합니다. MatrAIx는 세 가지 핵심 구성 요소로 이루어져 있습니다. 첫째, Persona 8B는 1,290개의 범주형 차원으로 표현된 83억 개의 가상 에이전트 기록을 포함합니다. 이러한 기록은 상관 관계가 있는 속성을 유지하는 의존 그래프에서 샘플링하거나 인간이 작성한 프로필에서 파생됩니다. 우리는 약 1백만 개의 고품질 코어셋을 공개하며, 여기에는 599,847개의 실제 데이터 기반 가상 에이전트와 400,000개의 합성 데이터 기반 가상 에이전트가 포함되어 있습니다. 둘째, MatrAIx Playground는 다양한 사용자가 디지털 제품을 평가하고 상호 작용할 수 있는 네 가지 환경(설문 조사, AI 챗봇, 웹, 앱)을 제공합니다. 셋째, MatrAIx는 전자 상거래, 소프트웨어, 금융 및 의료를 포함한 25개 이상의 분야에 걸쳐 1,010개의 애플리케이션 작업을 지원합니다. 우리는 여덟 가지 대표적인 작업에 대해 총 18,189회의 평가 실험을 수행했습니다. 가상 에이전트는 Claude Opus 4.8, GPT 5.5 및 Claude Haiku 4.5의 세 가지 LLM으로 구동되었습니다. 결과적으로 얻어진 피드백은 가격 인상 후 망설임, AI 비서가 실패했을 때 계속 진행할 의향, 지연 허용 범위 등 가상 에이전트의 배경에 따른 의사 결정 및 선호도의 다양성을 보여줍니다. 우리는 두 가지 주요 검증 연구를 수행했습니다. 첫째, 400회의 통제된 실험을 통해 10가지 행동 속성에 걸쳐 네 가지 환경에서 가상 에이전트의 일관성을 평가했습니다. 선언된 행동은 366회(91.5%) 동안 표현되었거나 정확하게 억제되었습니다. 둘째, 인간 및 LLM 평가자는 실제 데이터 기반 가상 에이전트 추출 품질을 평가했습니다. 전반적으로 MatrAIx는 다양한 시뮬레이션된 사용자를 통해 인공지능 시스템 및 디지털 제품을 평가하기 위한 엔드 투 엔드 인프라를 제공합니다.

Original Abstract

Human evaluation of AI systems and digital products is costly, slow, and difficult to scale. Offline evaluations are more scalable but often abstract away human diversity and interactive behavior. We therefore introduce MatrAIx, a population-scale simulated-user evaluation infrastructure for testing AI systems and digital products with heterogeneous users. MatrAIx has three core components: First, Persona 8B contains 8.3 billion persona records represented by 1,290 categorical dimensions. Records are either sampled from a dependency graph that preserves correlated attributes or derived from human-authored profiles. We release a quality-filtered coreset of approximately 1 million personas, comprising 599,847 human-grounded and 400,000 synthetic records. Second, the MatrAIx Playground provides four environments in which diverse users evaluate and interact with digital products: Survey, AI Chatbot, Web, and App. Third, MatrAIx provides 1,010 application tasks spanning more than 25 domains, including Commerce, Software, Finance, and Healthcare. We conducted 18,189 evaluation trials across eight representative tasks. Persona agents were powered by three LLMs: Claude Opus 4.8, GPT 5.5, and Claude Haiku 4.5. The resulting feedback captures how decisions and preferences vary across persona backgrounds, including hesitation after a price increase, willingness to continue after an AI assistant fails, and latency tolerance. We conducted two main validation studies: First, a 400-trial controlled study evaluated persona adherence across ten behavioral attributes and all four environments. The declared behavior was expressed or correctly suppressed in 366 trials (91.5%). Second, human and LLM judges evaluated the extraction quality of human-grounded personas. Overall, MatrAIx provides an end-to-end infrastructure for evaluating AI systems and digital products with diverse simulated human users.

0 Citations
0 Influential
30 Altmetric
150.0 Score
Original PDF

No Analysis Report Yet

This paper hasn't been analyzed by Gemini yet.

Log in to request an AI analysis.

댓글

댓글을 작성하려면 로그인하세요.

아직 댓글이 없습니다. 첫 번째 댓글을 남겨보세요!