Yatong Chen (陈雅桐)

Research Group Leader, Social Foundations of Computation
Max Planck Institute for Intelligent Systems, Tübingen

CV Google Scholar Twitter LinkedIn Email
Photo of Yatong Chen

About Me

I am a Research Group Leader in the Social Foundations of Computation department at the Max Planck Institute for Intelligent Systems in Tübingen, Germany, hosted by Moritz Hardt.

I build the theoretical foundations for incentive-aware evaluation. I study AI evaluation and algorithmic decision-making systems in which people, models, and institutions interact strategically. Currently, I am particularly interested in LLM evaluation ecosystems and in incentives in scientific discovery.

I received my Ph.D. in Computer Science and Engineering from the University of California, Santa Cruz, advised by Yang Liu, and was supported by the Chancellor’s Dissertation-Year Fellowship. In summer 2022 I was a student researcher at Google Brain, hosted by Ehsan Amid and Rohan Anil. Before that, I received an M.S. in Statistics from Stanford University (2019), advised by Johan Ugander, and a B.S. in Energy and Resources Engineering and B.A. in Economics from Peking University (2016).

Research

Leaderboards, drug-approval thresholds, and lending models all decide who gets ranked, approved, or rewarded, and the parties they evaluate adapt to them. I treat evaluation and decision-making as mechanisms: using game theory, mechanism design, statistics, and machine learning, I characterize how these rules shape behavior and design rules that prevent reward hacking and reward genuine improvement, under the constraints evaluators face in practice, such as imperfect verifiers and costly experiments. My work spans AI evaluation [1, 2], scientific discovery and regulation [3, 13], algorithmic recourse and lending [4, 6, 8], and fairness in deployed systems [9, 10, 11].

I’m especially excited about:

Incentive-aware AI evaluation

Benchmarks and leaderboards steer what model developers build, and developers in turn shape which benchmarks the field pays attention to. I want to understand this feedback loop and design evaluation protocols under which climbing a leaderboard requires real capability gains. For example, I study why benchmaxxing persists, and what alternative evaluation mechanisms can elicit model rankings that reflect true capabilities [1].

Relevant work on incentives and guarantees in AI evaluation

Incentive design for scientific discovery

Statistical tests act as contracts between those who produce evidence and those who act on it, such as regulators, journals, and funders. I study how significance thresholds and testing protocols determine which experiments get run, and how to design them so that evidence stays credible when firms and researchers respond strategically. For example, why might effective drugs for small patient populations, such as orphan drugs, never reach the market [3]?

Relevant work on strategic hypothesis testing and statistical contracts

Responsible decision-making under strategic adaptation

Algorithms increasingly decide who gets a loan, a job, or a public service, and the people affected respond to them. I design decision rules and learning algorithms that anticipate these responses and steer them toward genuine improvement rather than manipulation. Algorithmic recourse is one example: telling a rejected applicant what to change helps only if the institution has a reason to offer it and the advice leads to real improvement [4, 6, 8].

Relevant work on strategic classification, recourse, and performative prediction

Aligning algorithmic decisions with societal values

A decision system is aligned only if it keeps serving the values it was designed for after deployment, as the populations it serves change and the people it affects respond. I study how fairness guarantees hold up over time and across populations, and how to keep fair decision rules practical to deploy [9, 10, 11].

Relevant work on fairness under distribution shift and dynamics

News

Publications

* equal contribution

[1]
Leaderboard Incentives: Model Rankings under Strategic Post-Training
Yatong Chen, Guanhua Zhang, and Moritz Hardt.
ICML 2026.
[arXiv]
[2]
ROC-n-reroll: How Verifier Imperfection Affects Test-Time Scaling
Florian E. Dorner, Yatong Chen*, André F. Cruz*, and Fanny Yang.
ICLR 2026.
[arXiv]
[3]
Strategic Hypothesis Testing
Yatong Chen*, Safwan Hossain*, and Yiling Chen.
NeurIPS 2025. Spotlight, top 3%.
[arXiv]
[4]
To Give or Not to Give? The Impacts of Strategically Withheld Recourse
Yatong Chen, Andrew Estornell, Yevgeniy Vorobeychik, and Yang Liu.
AISTATS 2025. Preliminary version at the NeurIPS 2023 Workshop on Algorithmic Fairness through the Lens of Time.
[arXiv]
[5]
Performative Prediction with Bandit Feedback: Learning through Reparameterization
Yatong Chen, Wei Tang, Chien-Ju Ho, and Yang Liu.
ICML 2024.
[arXiv]
[6]
Learning to Incentivize Improvements from Strategic Agents
Yatong Chen, Jialu Wang, and Yang Liu.
Transactions on Machine Learning Research (TMLR). Preliminary version at the ICML 2021 Workshop on Algorithmic Recourse, Best Paper Award.
[paper] [code]
[7]
Model Transferability with Responsive Decision Subjects
Yatong Chen, Zeyu Tang, Kun Zhang, and Yang Liu.
ICML 2023. Preliminary version at the ICML 2022 Workshop on New Frontiers in Adversarial Machine Learning, Best Paper Award.
[paper] [slides] [code]
[8]
Incentivizing Recourse through Auditing in Strategic Classification
Andrew Estornell, Yatong Chen, Sanmay Das, Yang Liu, and Yevgeniy Vorobeychik.
IJCAI 2023.
[paper]
[9]
Tier Balancing: Towards Dynamic Fairness over Underlying Causal Factors
Zeyu Tang, Yatong Chen, Yang Liu, and Kun Zhang.
ICLR 2023. Preliminary version at the NeurIPS 2022 Workshop on Algorithmic Fairness through the Lens of Causality and Privacy.
[arXiv]
[10]
Fair Transferability Subject to Distribution Shift
Yatong Chen*, Reilly Raab*, Jialu Wang, and Yang Liu.
NeurIPS 2022. Preliminary version at the NeurIPS 2021 Workshop on Algorithmic Fairness through the Lens of Causality and Robustness.
[arXiv]
[11]
Metric-Fair Classifier Derandomization
Jimmy Wu, Yatong Chen, and Yang Liu.
ICML 2022. Spotlight.
[paper] [slides]
[12]
Decoupled Smoothing on Graphs
Alex Chin, Yatong Chen, Kristen M. Altenburger, and Johan Ugander.
WWW 2019. Oral, top 5%.
[paper] [code] [slides]

Workshop Papers

[13]
The Price of Sample-Size-Agnostic Contracts
Simone Garuffi, Jiduan Wu, and Yatong Chen.
NeurIPS 2026 Workshop on E-Values: From Statistics to ML.
[14]
Fast Implicit Constrained Optimization of Non-decomposable Objectives for Deep Networks
Yatong Chen, Abhishek Kumar, Yang Liu, and Ehsan Amid.
Has It Trained Yet? NeurIPS 2022 Workshop.
[paper]
[15]
Fishy: Layerwise Fisher Approximation for Higher-order Neural Network Optimization
Abel Peirson*, Ehsan Amid*, Yatong Chen, Vladimir Feinberg, Manfred K. Warmuth, and Rohan Anil.
Has It Trained Yet? NeurIPS 2022 Workshop.
[paper]
[16]
Decoupled Smoothing in Probabilistic Soft Logic
Yatong Chen*, Bryan Tor*, Eriq Augustine, and Lise Getoor.
15th International Workshop on Mining and Learning with Graphs (MLG), 2020.
[paper] [slides]

Distinctions

Selected Talks

Teaching & Service

Teaching

Service & Community

Contact

Email: yatong.chen [at] tuebingen.mpg.de
Max Planck Institute for Intelligent Systems, Tübingen, Germany
Google Scholar · Twitter · LinkedIn