leiyo@dtu.dk

Lei You (Ph.D.)

Assistant Professor

Technical University of Denmark (DTU)

I received my Ph.D. in Computer Science (specializing in Mathematical Optimization) from the Department of Information Technology at Uppsala University in 2019. During the PhD, I interned as a visiting data scientist at The Boston Consulting Group (BCG) Gamma. After the PhD, I worked as a data scientist in Bolt and Wolt (Doordash) in the domain of on-demand logistics optimization. I am a co-founder of https://cspaper.org.

Research Focus

My research develops counterfactual information design for trustworthy AI systems. The central question is what information must be exposed, preserved, or verified so that trust remains valid after a system changes? My work has evolved from optimization of coupled communication networks to information-theoretic auditing of coupled AI ecosystems: in both settings, reliability is not a property of isolated components, but of interactions under resource, information, and verification constraints. Methodologically, I combine information theory, counterfactual reasoning, causal inference, optimization, and AI systems to derive information laws for model replacement, verification bottlenecks, and system-level audit.

Research VISION

(See a full list of publications here)

Modern AI systems are compound systems of models, routers, retrievers, tools, verifiers, memories, and human review. Trustworthiness is therefore not only a property of one model. It is a property of information relations among components.

My research develops information laws for trustworthy compound AI systems:

What can be replaced?
What must be verified?
What information must remain available for the system to stay trustworthy?

Area 1: Trustworthy Compound AI Systems

I study AI systems as ecosystems of interacting components. The central question is not only whether one model performs well, but whether its behavior can be reproduced, repaired, or verified by the remaining system under explicit information constraints. This line includes matched in-silico quasi-experimental design, DISCO/PIER audits, and Minimum Viable Replacement: the information cost of replacing a blocked model.

Selected Publications

L. You ✉, "Quantifying Model Uniqueness in Heterogeneous AI Ecosystems", preprint. [OpenPrint] [code]
L. You ✉, L. Cao, M. Nilsson, B. Zhao, and L. Lei, "Distributional Counterfactual Explanation With Optimal Transport", International Conference on Artificial Intelligence and Statistics (AISTATS) 2025 (Oral, top 2%). [arXiv] [code]
L. You ✉ and H. V. Cheng ✉, "SWAP: Sparse Entropic Wasserstein Regression for Robust Network Pruning", International Conference on Learning Representations (ICLR) 2024. [arXiv] [code]

Area 2: Verification-Limited Trustworthy AI

AI systems can now generate candidates, reviews, claims, code patches, and hypotheses at scale. The bottleneck is verification. This line studies how scarce human or automated checks turn untrusted candidate information into reliable knowledge, with applications to peer review, scientific assessment, RAG systems, and agentic workflows. I co-founded CSPaper (https://cspaper.org) as an open research infrastructure and a testbed for studying verification-limited AI systems.

Selected Publications

L. You ✉, "Epistemic Throughput: Fundamental Limits of Attention-Constrained Inference", preprint. [OpenPrint] [code]
L. You ✉, L. Cao, and I. Gurevych, "Preventing the Collapse of Peer Review Requires Verification-First AI", preprint. [OpenPrint]
C. Yu, T. Shi, V. Uotila, S. Deng, L. You, and B. Zhao, "Vista: Verifier-in-the-Loop Agentic RL for Semantic Program Synthesis in Quantum Computing", ACM Conference on AI and Agentic Systems (CAIS) 2026. [link]

Area 3: Counterfactuals and Trustworthy Learning

This line develops counterfactual explanation methods that respect data geometry, distributional structure, and fairness constraints. It provides the methodological basis for my broader programme: counterfactual claims become scientifically useful only when the admissible transformation and its cost are made explicit.

Selected Publications

L. You ✉, Y. Bian, and L. Cao, "Joint Distribution–Informed Shapley Values for Sparse Counterfactual Explanations", International Conference on Learning Representations (ICLR) 2026 [arXiv] [code] [software].
L. Zhu, Y. Bian, and L. You ✉, "FairSHAP: Preprocessing for Fairness Through Attribution-Based Data Augmentation", International Conference on Artificial Intelligence and Statistics (AISTATS) 2026. [OpenPrint] [code]
Y. Gu, L. Cao, B. Zhao, L. Lei, and L. You ✉, "DISCOVER: A Solver for Distributional Counterfactual Explanations", European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD) 2026. [arXiv] [code]

Page updated

Google Sites

Report abuse