Chengshuai Shi
 |
Postdoctoral Fellow Princeton Language and Intellegence, Princeton University Email: cs1083 at princeton dot edu Phone: 434-218-9860 |
I am currently a Postdoctoral Fellow at the Princeton Language and Intelligence (PLI) initiative at Princeton University, working with Professor Chi Jin. Prior to joining PLI, I worked for a year as a Senior Machine Learning Research Engineer in the AI group at Bloomberg, New York City.
I received my Ph.D. in Electrical Engineering from the University of Virginia in 2024, where I was advised by Professor Cong Shen. During my Ph.D. (2021–2024), I was honored to be supported by the Bloomberg Data Science Ph.D. fellowship.
My research interests lie in machine learning, with a focus on intelligent decision-making. Specifically, I work on:
Foundational principles in areas such as reinforcement learning, multi-armed bandits, game theory, and multi-agent systems;
Applications in emerging fields including wireless communication, recommender systems, and large language models
News
09/2025: I have joined Princeton University as a Postdoctoral Fellow at the Princeton Language and Intelligence (PLI) initiative! Feeling excited for this new adventure!
05/2025: One paper accepted to UAI 2025!
“Augmenting Online RL with Offline Data is All You Need: A Unified Hybrid RL Algorithm Design and Analysis”: We propose a unified RL approach that combines offline data with online exploration, yielding improved sub-optimality and regret bounds while revealing distinct coverage requirements. This is a joint work with Ruiquan Huang (PSU), Donghao Li (UVA), Prof. Cong Shen (UVA), and Prof. Jing Yang (UVA).
01/2025: One paper accepted to ICLR 2025!
“Building Math Agents with Multi-Turn Iterative Preference Learning”: We propose a multi-turn direct preference learning framework to enhance the mathematical reasoning capabilities of large language models (LLMs), with its superiority demonstrated by both theoretical guarantees and empirical results. This is a joint work with many amazing researchers from UIUC, Princeton, Google Research, and Google Deepmind.
01/2025: One paper accepted to AISTATS 2025!
“Cost-Aware Optimal Pairwise Pure Exploration”: We extend the canonical best-arm identification (BAI) study in two directions: (1) tackling a broader set of tasks, termed pairwise exploration tasks, covering BAI, ranking identification, top-k identification and others in a unified manner; (2) minimizing the arm-specific costs, instead of total number of arm-pulling. This is a joint work with Di Wu (UVA), Ruida Zhou (UCLA), and Prof. Cong Shen (UVA).
09/2024: Three papers accepted to NeurIPS 2024!
[Co-first-authored] “Efficient Prompt Optimization Through the Lens of Best Arm Identification”: A systematical relationship between prompt optimization and best arm identification in MAB is established in this work. Built upon this relationship, a general framework is proposed to efficiently perform prompt optimization. This is a joint work with Kun Yang (UVA), Zihan Chen (UVA), Prof. Jundong Li (UVA), Prof. Jing Yang (PSU), and Prof. Cong Shen (UVA).
A preliminary version appears in ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, May 2024.
[First-authored] “Transformers as Game Players: Provable In-context Game-playing Capabilities of Pre-trained Models”: We provide theoretical understanding and empirical evidences on the in-context learning capabilities of pre-trained transformers in multiplayer competitive games. This is a joint work with Kun Yang (UVA), Prof. Jing Yang (PSU), and Prof. Cong Shen (UVA).
A preliminary version appears in ICLR 2024 Workshop on Generative Models for Decision Making, May 2024.
“Mixture of Demonstrations for In-Context Learning”: We provide a novel framework to effectively select demosntrations for prompts used in in-context learning, leveraing an expert-wise training strategy. This is a joint work with Song Wang (UVA), Zihan Chen (UVA), Prof. Cong Shen (UVA), and Prof. Jundong Li (UVA).