Research

Working Papers

  • Learning to Design Information (with Tao Lin), Job Market Paper. [pdf]

    • Talks: INFORMS 2025 Annual Meeting, Conference on Information Systems and Technology (CIST), invited talk at Institute of Theoretical Computer Science of Shanghai University of Finance and Economics, invited talk at School of Data Science at Fudan University

    Information designers, such as large language models and online platforms, often do not know the subjective beliefs of their receivers or users. We construct learning algorithms enabling the designer to learn the receiver’s belief through repeated interactions. Our learning algorithms are robust to the receiver’s strategic manipulation of the learning process of the designer. We study regret relative to two benchmarks to measure the performance of the algorithms. The static benchmark is T times the single-period optimum for the designer under a known belief. The dynamic benchmark, which is stronger, characterizes global dynamic optimality for the designer under a known belief. Our learning algorithms allow the designer to achieve no regret against both benchmarks at fast speeds of O(1/T log^2 T), significantly faster than other approaches in the literature.

  • Information Design with Unknown Prior (with Tao Lin), latest draft: Oct 2025 [pdf] (under review at Theoretical Economics)

    • Acceptance: 16th Innovations in Theoretical Computer Science Conference (ITCS 2025)
    • Talks: World Congress of the Econometric Society, Stony Brook Game Theory Conference, Econometric Society Interdisciplinary Frontiers (ESIF) conference on Economics and AI+ML, World Congress of the Game Theory Society, 2025 North American Winter Meeting of the Econometric Society

    Information designers, such as online platforms, often do not know the beliefs of their receivers. We design learning algorithms so that the information designer can learn the receivers’ prior belief from their actions through repeated interactions. Our learning algorithms achieve no regret relative to the optimality for the known prior at a fast speed, achieving a tight regret bound Θ(log T) in general and a tight regret bound Θ(log log T) in the important special case of binary actions.

  • From Best Responses to Learning: Investment Efficiency in Dynamic Environment (with Qianfan Zhang and Weiqiang Zheng), latest draft: Jun 2025 [pdf]

    • Talks: Stony Brook Game Theory Conference

    We study the social welfare of a mechanism in a dynamic environment where a learning investor can make a costly investment to change her value. In many real-world problems, the common assumption that the investor could always make the best responses, i.e., choosing her utility-maximizing investment option, is unrealistic due to incomplete information in a dynamically evolving environment. To address this challenge, we instead consider an investor who makes her investment decisions using no-regret online learning algorithms in a dynamic environment. We study the approximation ratio, as the measurement of the performance of an allocation algorithm, in the dynamic learning environment against different benchmarks. First, we show that the approximation ratio in the static environment remains unchanged in the dynamic environment for the best-in-hindsight benchmark. Second, we characterize matching lower and upper bounds of the approximation ratio for a time-varying benchmark.

  • Investment Incentives with Limited Cognition, latest draft: Oct 2023 [pdf]

    • Talks: Stony Brook Game Theory Conference

    A bidder may have an investment incentive, i.e., to increase his value with a cost to obtain the allocation of an item, but that does not guarantee to benefit himself or social welfare, mainly due to the computational infeasibility of resource allocation problems and the bidder’s inability to make his best responses in auctions. In this paper, I study the investment incentives of a bidder, who cannot make his best responses due to limited cognition, in strategy-proof mechanisms.

    For a single auction, without knowing the price, a bidder makes his response, which includes his investment decision, based on his prediction of his allocation result. I show that when the prediction is correct, the allocation algorithm needs to exclude confirming negative externalities (a property proposed by Akbarpour et al., 2023) to achieve approximately optimal welfare. When the prediction is incorrect, I characterize a stronger property, downward preserving positive externalities (DPONE), that an allocation algorithm should satisfy to maintain near-optimal welfare.

    For repeated auctions, consider a cognitively limited bidder who uses an online learning algorithm to guide his decision-making. I propose a measure, potential, that links regret with Pareto efficiency to measure how reliable a simple strategy is as a benchmark for the regret of a cognitively limited bidder. I show that the value of the potential increases with the frequency of the preservation of near-optimal welfare in a privately investment-friendly environment. The approach also provides other new insights into the dynamic impact of the investment incentives of a cognitively limited bidder, including the implication of a privately investment-friendly environment on itself as being socially investment-friendly but not vice versa.

  • Principal-agent LLM Alignment (with Lingkai Kong, Tao Lin, Milind Tambe, and Haichuan Wang), draft coming soon!

    We study inference-time alignment of large language models (LLMs) when users may strategically misreport their preferences. Existing inference-time methods for LLMs, such as exponential tilting by a user-provided reward, are low-cost and flexible. But those methods allow strategic manipulation: users may be incentivized to misreport their preferences, which leaves feedback systematically biased. We model this interaction in a principal–agent framework, where a service provider (principal) steers the LLM policy (agent) by choosing a reward-shaping contract subject to a shaping budget and the soft KL constraint induced by tilting. Within this framework, we formulate a bilevel optimization problem and derive the optimal shaping rule. We show that the resulting principal–agent alignment mechanism is incentive compatible: users maximize expected utility by truthfully reporting their reward. We then give a Monte Carlo estimator and stochastic gradient procedure that implements the shaping using only samples from the baseline model and the reported reward. That yields a practical and fine-tuning-free alignment method that improves personalization and preference satisfaction while preserving the strengths of a pretrained baseline.

  • Repeated Blackwell Experiments (with Tao Lin and Qianfan Zhang), draft coming soon!

    We study how an information designer reuses a signaling scheme over time, rather than redesigning it in each interaction. The designer cares about her long-run average payoff under repetition compared to the payoff of an optimal static experiment. This yields an online analogue of Blackwell’s comparison in large samples and regret-style benchmarks for repeated experiments under minimal assumptions on the state process.

Publication

  • Information Design with Unknown Prior (Extended Abstract) [pdf]

    Tao Lin, Ce Li

    16th Innovations in Theoretical Computer Science Conference, ITCS 2025 Vol. 325. LIPIcs. Schloss Dagstuhl-Leibniz-Zentrum fur Informatik, 2025, 72:1-72:1. doi: 10.4230/LIPICS.ITCS.2025.72.

  • Association of degree of loss aversion and grey matter volume in superior frontal gyrus by voxel-based morphometry [pdf]

    Ce Li, Xueqin Wang, Canhong Wen, Haizhu Tan

    Brain Imaging Behavior. 2018 Feb; 14(1):89-99. PMID: 30328557. doi: 10.1007/s11682-018-9962-5.

Conference Papers

  • Spatial-temporal Modeling of Ambient PM 2.5 Elemental Concentrations in Eastern Massachusetts [poster]

    Ce Li, Jeremiah Zhe Liu, Weeberb Requia, Choong-Min Kang, Michael Wolfson, Joel Schwartz, Petros Koutrakis, Brent A. Coull.

    2019 Harvard & MIT Air, Climate & Energy Center Science Advisory Committee Meeting

  • Spatial Multiresolution Analysis of PM 2.5 and Mortality in New England [poster]

    Ce Li, Joseph Antonelli, Weeberb Requia, Antonella Zanobetti, Itai Kloog, Petros Koutrakis, Joel Schwartz, Brent A. Coull.

    2019 Harvard & MIT Air, Climate & Energy Center Science Advisory Committee Meeting