Tianyi Peng

Tianyi Peng

Assistant Professor at Columbia University

Decision, Risk, and Operations Division, Columbia Business School

Member of Columbia Data Science Institute

Member of the Data, Agents, and Processes Lab (DAPLab)

Member of the Columbia AI Agents Initiative

Email: tianyi.peng [at] columbia [dot] edu

About Me

I am interested in preparing the world for human coexistence with artificial superintelligence. My research spans:

  • Optimizing large-scale decision-making systems and AI agents through reinforcement learning (RL), ensuring continuous alignment with business and societal objectives.
  • Leveraging AI to simulate human and system behavior ("digital twin", or "environments"), enabling scalable synthetic data for experimentation, personalization, world modeling, and RL training.
  • Reducing the cost and latency of deploying AI agents through system- and algorithm-level optimizations.

My research uncovers fundamental connections between RL and causal inference (e.g., Markovian Interference in Experiments), as well as between RL and quantum entanglement (e.g., Multi-Agent Markov Entanglement). I also work closely with industry partners to deploy AI in real-world settings, including online social platforms, e-commerce, retail, and finance (e.g., Cursor, TikTok, Anheuser-Busch InBev, Liberty Mutual).


Collectively, these studies have received recognition from the academic community, including the Daniel H. Wagner Prize, the INFORMS Junior Faculty Interest Group Paper Prize, the Applied Probability Society Best Student Paper Prize, and the Revenue Management and Pricing Jeff McGill Student Paper Award.


I believe in "Mens et Manus": new ideas emerge from hands-on building. Before joining Columbia, I was a founding member of Cimulate.AI, where we built CommerceGPT from scratch: an end-to-end system for e-commerce search based on transformers and RL. Earlier, I was a competitive programmer, ranking 8th in China's Team Selection Contest for the International Olympiad in Informatics (IOI) in 2013.


I received my Ph.D. from MIT in 2023, advised by Vivek Farias, and a Bachelor's degree from Yao Class at Tsinghua University in 2017.

Selected Publications

(* indicates alphabetical ordering of authors, a celebrated tradition in Operations Research)

Speculative Actions: A Lossless Framework for Faster Agentic Systems

Naimeng Ye, Arnav Ahuja, Georgios Liargkovas, Yunan Lu, Kostis Kaffes, Tianyi Peng

Agentic systems are often slow due to their inherently sequential action execution. Inspired by speculative decoding, we introduce speculative execution and caching to enable action-level parallelism, providing a new lever for reducing latency in agentic workflows.

Deploying at Cursor

Multi-agent Markov Entanglement

* Shuze Chen, Tianyi Peng

NeurIPS 2025, Spotlight (~3%)

Winner, INFORMS Junior Faculty Interest Group (JFIG) Paper Competition

Second Place, INFORMS George Nicholson Student Paper Competition

We discovered that the effectiveness of value decomposition in reinforcement learning hinges on an intrinsic "entanglement" structure within transition matrices, a relationship that mirrors the mathematical form of quantum entanglement.

Twin-2K-500: A dataset for building digital twins of over 2,000 people based on their answers to over 500 questions

Olivier Toubia, George Z. Gui, Tianyi Peng, Daniel J. Merlau, Ang Li, and Haozhe Chen

Marketing Science

[Twin-2K-500 Dataset] [GitHub Documentation]

One of the first large-scale open-source benchmarks for LLM-based digital-twin simulations, spanning more than 2,000 individuals and over 500 questions. The dataset includes a comprehensive suite of demographic, psychological, economic, personality, and cognitive measures, as well as replications of behavioral-economics experiments and a pricing survey. These efforts have been featured by Columbia Business School, Prolific, and Harvard Business Review.

Tail-Optimized Caching for LLM Inference

Wenxin Zhang, Yueying Li, Ciamac C. Moallemi, Tianyi Peng

NeurIPS 2025

How Well do LLMs Compress their Own Chain-of-Thought? A Token Complexity Approach

Ayeong Lee, Ethan Che, Tianyi Peng

Efficient Systems for Foundation Model, ICML 2025 Workshop

We conduct a benchmark test across diverse prompting strategies for reasoning. Remarkably, their accuracy-length tradeoffs fall on a roughly similar curve, despite the strategies being very different. Furthermore, the empirical results suggest a "token complexity" phenomenon: each question has a minimal number of tokens required for successful reasoning.

Picard Iteration: Massive Speedups in Policy Simulation for Supply Chain and General RL

* Vivek F. Farias, Joren Gijsbrechts, Aryan Khojandi, Tianyi Peng, Andrew Zheng

ICML 2025

LLM Generated Persona is a Promise with a Catch

Ang Li, Haozhe Chen, Hongseok Namkoong, Tianyi Peng

NeurIPS 2025, Position Paper

LLM-generated personas are already widely used across industries. We uncover a notable "blue-shift" phenomenon: as more LLM-generated persona attributes are provided to the model, its responses become increasingly progressive, eventually shifting simulated U.S. presidential voting outcomes toward one party across all states. For a short summary, see this post.

Markovian Interference in Experiments

* Vivek F. Farias, Andrew A. Li, Tianyi Peng, Andrew Zheng

Major Revision in Management Science

Winner, Applied Probability Society (APS) Best Student Paper Prize 2022

Winner, Jeff McGill Student Paper Award 2022

NeurIPS 2022, Oral (~2%)

[Video in Chinese]

Worked with Douyin for deployment (RecSys 2023)

Learning Treatment Effects in Panels with General Intervention Patterns

* Vivek F. Farias, Andrew A. Li, Tianyi Peng

Finalist, MSOM Best Student Paper Prize 2022

NeurIPS 2021, Oral (~0.6%)

[A Python Package for Causal Inference in Panels]

We deployed this work with Anheuser-Busch InBev, resulting in an innovative experimentation platform that is impacting hundreds of millions of dollars per month (INFORMS news).

Generalized Synthetic Control for TestOps at ABI: Models, Algorithms, and Infrastructure

* Luis Costa, Vivek F. Farias, Patricio Foncea, Jingyuan (Donna) Gan, Ayush Garg, Ivo Rosa Montenegro, Kumarjit Pathak, Tianyi Peng, Dusan Popovic

INFORMS Journal on Applied Analytics

Winner, Daniel H. Wagner Prize 2022

Simulating Large Quantum Circuits on a Small Quantum Computer

Tianyi Peng, Maris Ozols, Aram Harrow, Xiaodi Wu

Physical Review Letters 125, 150504 (2020) (PRL)

This schema has been implemented by IBM Quantum (Nature 2024).

Optimal Remote Entanglement Distribution

Wenhan Dai, Tianyi Peng, Moe Win

IEEE Journal on Selected Areas in Communications (IEEE-JSAC), vol. 38, no. 3, pp. 540–556, 2020

Best Paper Award, International Conference on Computing, Networking and Communications (ICNC 2020)

Publications ▼

(* indicates alphabetical ordering of authors, a celebrated tradition in Operations Research)

RAISE: Reliable Agent Improvement via Simulated Experience

Sahar Omidi Shayegan, Joshua Meyer, Victor Shih, Sebastian Sosa, Tianyi Peng, Kostis Kaffes, Eugene Wu, Andi Partovi, Mehdi Jamei

Scaling Environments for Agents Workshop, NeurIPS 2025 Workshop

Multi-agent Markov Entanglement

* Shuze Chen, Tianyi Peng

NeurIPS 2025, Spotlight (~3%)

Winner, INFORMS Junior Faculty Interest Group (JFIG) Paper Competition

Second Place, INFORMS George Nicholson Student Paper Competition

We discovered that the effectiveness of value decomposition in reinforcement learning hinges on an intrinsic "entanglement" structure within transition matrices, a relationship that mirrors the mathematical form of quantum entanglement.

Picard Iteration: Massive Speedups in Policy Simulation for Supply Chain and General RL

* Vivek F. Farias, Joren Gijsbrechts, Aryan Khojandi, Tianyi Peng, Andrew Zheng

ICML 2025

Differences-in-Neighbors for Network Interference in Experiments

* Tianyi Peng, Naimeng Ye, Andrew Zheng

EC 2025

Finalist, INFORMS Jeff McGill Student Paper Award 2025

Speculative Actions: A Lossless Framework for Faster Agentic Systems

Naimeng Ye, Arnav Ahuja, Georgios Liargkovas, Yunan Lu, Kostis Kaffes, Tianyi Peng

Agentic systems are often slow due to their inherently sequential action execution. Inspired by speculative decoding, we introduce speculative execution and caching to enable action-level parallelism, providing a new lever for reducing latency in agentic workflows.

Deploying at Cursor

Tail-Optimized Caching for LLM Inference

Wenxin Zhang, Yueying Li, Ciamac C. Moallemi, Tianyi Peng

NeurIPS 2025

How Well do LLMs Compress their Own Chain-of-Thought? A Token Complexity Approach

Ayeong Lee, Ethan Che, Tianyi Peng

Efficient Systems for Foundation Model, ICML 2025 Workshop

We conduct a benchmark test across diverse prompting strategies for reasoning. Remarkably, their accuracy-length tradeoffs fall on a roughly similar curve, despite the strategies being very different. Furthermore, the empirical results suggest a "token complexity" phenomenon: each question has a minimal number of tokens required for successful reasoning.

Throughput-Optimal Scheduling Algorithms for LLM Inference and AI Agents

Yueying Li, Tianze Deng, Jim Dai*, Tianyi Peng*

Data Mixture Optimization: A Multi-Fidelity Multi-Scale Bayesian Framework

Tzu-Ching Yen, Andrew Wei Tung Siah, Haozhe Chen, C. Daniel Guetta, Tianyi Peng, Hongseok Namkoong

NeurIPS 2025

Twin-2K-500: A dataset for building digital twins of over 2,000 people based on their answers to over 500 questions

Olivier Toubia, George Z. Gui, Tianyi Peng, Daniel J. Merlau, Ang Li, and Haozhe Chen

Marketing Science

[Twin-2K-500 Dataset] [GitHub Documentation]

One of the first large-scale open-source benchmarks for LLM-based digital-twin simulations, spanning more than 2,000 individuals and over 500 questions. The dataset includes a comprehensive suite of demographic, psychological, economic, personality, and cognitive measures, as well as replications of behavioral-economics experiments and a pricing survey. These efforts have been featured by Columbia Business School, Prolific, and Harvard Business Review.

LLM Generated Persona is a Promise with a Catch

Ang Li, Haozhe Chen, Hongseok Namkoong, Tianyi Peng

NeurIPS 2025, Position Paper

LLM-generated personas are already widely used across industries. We uncover a notable "blue-shift" phenomenon: as more LLM-generated persona attributes are provided to the model, its responses become increasingly progressive, eventually shifting simulated U.S. presidential voting outcomes toward one party across all states. For a short summary, see this post.

Digital Twins as Funhouse Mirrors: Five Key Distortions

Tianyi Peng, George Gui, Daniel J. Merlau, Grace Jiarui Fan, Malek Ben Sliman, Melanie Brucks, Eric J. Johnson, Vicki Morwitz, Abdullah Althenayyan, Silvia Bellezza, Dante Donati, Hortense Fong, Elizabeth Friedman, Ariana Guevara, Mohamed Hussein, Kinshuk Jerath, Bruce Kogut, Akshit Kumar, Kristen Lane, Hannah Li, Patryk Perkowski, Oded Netzer, Olivier Toubia

In Submission

What if we replaced human participants with their digital-twin simulations in survey research today? Across 19 pre-registered studies conducted by an interdisciplinary team at Columbia Business School, we find five systematic distortions that researchers should be mindful of when using LLM-based synthetic personas.

AI Agents for Web Testing: A Case Study in the Wild

Naimeng Ye, Xiao Yu, Ruize Xu, Tianyi Peng, Zhou Yu

NeurIPS 2025 Workshop on Bridging Language, Agent, and World Models

E-GEO: A Testbed for Generative Engine Optimization in E-Commerce

* Puneet S Bagga, Vivek F Farias, Tamar Korkotashvili, Tianyi Peng, Yuhang Wu

Estimation of Treatment Effects Under Nonstationarity via the Truncated Policy Gradient Estimator

* Ramesh Johari, Tianyi Peng, Wenqian Xing

arXiv preprint arXiv:2506.05308

Synthetic Control Approach to Digital Twin Calibration

* Grace Jiarui Fan, Chengpiao Huang, Tianyi Peng, Kaizheng Wang, Yuhang Wu

A preliminary project won Disney Data & Analytics Women's Award 2025

Enterprise Digital Twin: A Case Study for Enron

Grace Jiarui Fan, Xiaotong Tang, Glen Park, Vivian Zhang, Tianyi Peng

QGym: Scalable Simulation and Benchmarking of Queuing Network Controllers

Haozhe Chen, Ang Li, Ethan Che, Tianyi Peng, Jing Dong, Hongseok Namkoong

NeurIPS 2024, Dataset and Benchmarks Track

[Code]

Performance of LLMs on Stochastic Modeling Operations Research Problems: From Theory to Practice

* Akshit Kumar, Tianyi Peng, Yuhang Wu, Assaf Zeevi

Winter Simulation Conference 2024

Correcting for Interference in Experiments: A Case Study at Douyin

* Vivek F. Farias, Hao Li, Tianyi Peng, Xinyuyang Ren, Huawei Zhang, Andrew Zheng

RecSys 2023

Markovian Interference in Experiments

* Vivek F. Farias, Andrew A. Li, Tianyi Peng, Andrew Zheng

Major Revision in Management Science

Winner, Applied Probability Society (APS) Best Student Paper Prize 2022

Winner, Jeff McGill Student Paper Award 2022

NeurIPS 2022, Oral (~2%)

[Video in Chinese]

Worked with Douyin for deployment (RecSys 2023)

Generalized Synthetic Control for TestOps at ABI: Models, Algorithms, and Infrastructure

* Luis Costa, Vivek F. Farias, Patricio Foncea, Jingyuan (Donna) Gan, Ayush Garg, Ivo Rosa Montenegro, Kumarjit Pathak, Tianyi Peng, Dusan Popovic

INFORMS Journal on Applied Analytics

Winner, Daniel H. Wagner Prize 2022

Synthetically Controlled Bandits

* Vivek F. Farias, Ciamac C. Moallemi, Tianyi Peng, Andrew Zheng

MSOM Service Management SIG 2022; Major Revision in Management Science

Uncertainty Quantification for Low-Rank Matrix Completion with Heterogeneous and Sub-Exponential Noise

* Vivek F. Farias, Andrew Li, Tianyi Peng

AISTATS 2022

Learning Treatment Effects in Panels with General Intervention Patterns

* Vivek F. Farias, Andrew A. Li, Tianyi Peng

Finalist, MSOM Best Student Paper Prize 2022

NeurIPS 2021, Oral (~0.6%)

[A Python Package for Causal Inference in Panels]

We deployed this work with Anheuser-Busch InBev, resulting in an innovative experimentation platform that is impacting hundreds of millions of dollars per month (INFORMS news).

Fixing Inventory Inaccuracies at Scale

* Vivek F. Farias, Andrew Li, Tianyi Peng

ICML 2021; Manufacturing & Service Operations Management; MSOM Supply Chain SIG 2022

The Limits to Learning a Diffusion Model

* Jackie Baek, Vivek F. Farias, Andreea Georgescu, Retsef Levi, Tianyi Peng, Deeksha Sinha, Joshua Wilde, Andrew Zheng

EC 2021, Management Science

Simulating Large Quantum Circuits on a Small Quantum Computer

Tianyi Peng, Maris Ozols, Aram Harrow, Xiaodi Wu

Physical Review Letters 125, 150504 (2020) (PRL)

This schema has been implemented by IBM Quantum (Nature 2024).

Quantum Queuing Delay

Wenhan Dai, Tianyi Peng, Moe Win

IEEE Journal on Selected Areas in Communications (IEEE-JSAC), vol. 38, no. 3, pp. 605–618, 2020

Preliminary version: Queuing Delay for Quantum Networks, International Conference on Computing, Networking and Communications (ICNC 2020)

Optimal Remote Entanglement Distribution

Wenhan Dai, Tianyi Peng, Moe Win

IEEE Journal on Selected Areas in Communications (IEEE-JSAC), vol. 38, no. 3, pp. 540–556, 2020

Best Paper Award, International Conference on Computing, Networking and Communications (ICNC 2020)

Efficient and Robust Physical Layer Key Generation

Tianyi Peng, Wenhan Dai, Moe Win

Military Communications Conference 2019 (MILCOM 2019)

Remote State Preparation for Multiple Parties

Wenhan Dai, Tianyi Peng, Moe Win

IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2019), Invited Paper

Quantum Uncertainty Relation of Coherence

Xiao Yuan, Ge Bai, Tianyi Peng, Xiongfeng Ma

Physical Review A 96 (3), 032313 (2017)

Tight Detection Efficiency Bounds of Bell Tests In No-signaling Theories

Zhu Cao, Tianyi Peng (co-first author)

Physical Review A 94, 042126 (2016)