280-character bio
Andreas Haupt is an AI Institute Fellow-in-Residence at Schmidt Sciences and a fellow at Stanford's Digital Economy Lab and Harvard Kennedy School. He studies how AI systems represent and are evaluated against diverse human preferences, publishing in NeurIPS, ACM EC, and the AER.
Tagline
Use AI to Learn about People's Preferences
Media AssetsBANNERFORMAL HEADSHOTINFORMAL HEADSHOTPANORAMAWHITEBOARD HEADSHOT
Publications
A full publication list is available as pdf and BibTeX, and a more complete list on Google Scholar. ‡ indicates equal contribution or alphabetic author listing, † equal supervision. Only published work is shown by default; select another type, or All types, to see the rest.
SimulacraBench
Organizers: A. Haupt
Conference on Neural Information Processing Systems, Sydney, 2026.
The optimal weight of an item rises with its welfare relevance and improvability and falls with its variance. Improvability is estimated from what a 10% improvement costs on the EvoLM 4B suite; the table shows one OLMES item per improvability quantile from the paper's Table 1.
A benchmark is reported as a measurement but works as an incentive: once labs optimize against it, the weight on each item steers effort. We model benchmark aggregation as a multitask principal–agent problem. The optimal item weights have a closed form, v* = (M + rΣ)−1Mw—higher for items that matter for welfare (w) and that effort can improve (M), lower for noisy items (Σ) when the lab is risk-averse (r). On the OLMES suite, with welfare weights from workers' own task ratings, all 52 mathematics items are Pareto-dominated by general-knowledge items under a pro-worker welfare definition, and only ~6% of the suite's "automation loading" falls on activities workers want automated. Evaluation designers should publish why each item carries the weight it does.
A length penalty lowers the length bias, but the optimization pressure resurfaces as overconfidence (schematic).
Fixing one bias in a reward model (length, sycophancy, style) can simply redirect optimization pressure elsewhere. We prove that success, substitution, and overcorrection are indistinguishable under the audit distributions on which mitigations are usually reported—even with an oracle true reward. Empirically, a length penalty under GRPO shifts pressure into overconfidence with lower factual accuracy, and a published length-debiasing operator's bias reappears under best-of-n sampling for three of four state-of-the-art reward models. Prescription: evaluate under the policy-induced distribution and track several biases at once.
Position: AI Development Should Prioritize Cognitive Security
B. El, S. Su, A. Pappu, P. Yin, J. Heng, E. Heng, R.Z. Wang, A. Haupt†, J. Zou†
Advances in Neural Information Processing Systems (Position Paper), 2026.
From worker interviews to an occupational task inventory, with the counts from the six-occupation pilot.
Every estimate of how AI affects jobs rests on an inventory of what workers actually do—and the public inventory, O*NET, is updated slowly. TaskNet interviews incumbent workers with a mixed-initiative language-model interviewer, standardizes their reports into task statements, and aggregates them into occupational inventories. In six occupations, 132 workers produced 2,726 reports that became 202 task statements; O*NET analysts accepted all of them (versus ~65% of write-in suggestions under the standard process); 318 further incumbents validated them with 7,923 ratings. The inventory covers 89% of held-out workers' tasks (O*NET: 79%) and adds 79 tasks missing from the published list. AI-exposure estimates shift in both directions—upward for software developers and lawyers. Built with the National Center for O*NET Development and Mercor; a 250-occupation scale-up is in preparation.
Contextually Private Mechanisms
A. Haupt‡, Z. Hitzig‡
American Economic Review 116 (9), p. 3223-3262 (Lead Article), 2026.
The corners lemma: when two agents are jointly but not individually pivotal, whoever is asked first is told apart from a profile with the same outcome, so no protocol implementing the rule is contextually private.
Mechanism design usually asks what to implement and models the mechanism as a one-shot report of types. This paper studies the extensive form instead: a protocol asks participants questions one at a time, and what it learns depends on which question it asks, of whom, and when. A protocol violates contextual privacy when it tells two people apart who would have received the same outcome anyway. Violations are unavoidable exactly when agents are collectively, but not individually, pivotal. Beyond that, the order of questions is a design choice: delaying a question until its answer is needed for the outcome means fewer people are told apart for no reason, so delay improves privacy, and different extensive forms of the same allocation rule can differ sharply in how much they learn. The paper connects auction theory to Helen Nissenbaum's contextual integrity: "information not needed for the outcome" becomes a theorem about protocols.
AISPA: User-Centric System Prompt Auditing for Large Language Model Applications
X. Lin, S. Zhu, S. Yang, Z. Zhang, H. Zhang, Y. Zhao, C. Qian, T. Wang, Z. Zhang, Z. Yuan, D. Wu, J. Wu, Y. Si, J. Liu, B. Bi, R. Mahari, T. South, D. Greenwood, Z. He, R. Bommasani, S. Kazinnik, A. Haupt, S. Marro, E. Brynjolfsson, A. Pentland, J. Pei
Philosophers' positions across questions are dispersed; the models' positions line up, manufacturing consensus where the profession disagrees (schematic).
Can a language model stand in for an expert panel? We tested seven models against 277 professional philosophers and the PhilPapers survey (N = 1,785). Models reproduce individual positions tolerably, but they over-correlate judgments across questions, manufacturing consensus where the profession disagrees—partly a "specialist effect": the model assumes specialists in a domain share views. For anyone using "silicon samples" as proxies for publics, the heterogeneity collapses exactly where it matters.
Discussion of "The ICML 2023 Ranking Experiment" by Su et al.
A. Haupt, S. Koyejo
Journal of the American Statistical Association 121 (554), p. 853-854, 2026.
PublishedAI EvaluationAggregation
General Preference Reinforcement Learning
M. Umer, M.A. Mohsin, A. Bilal, A. Chaudhry, A. Haupt, S. Koyejo, E. Fox, J.M. Cioffi
A centaur evaluation scores the uplift of a human–AI team over either alone, instead of the AI by itself.
Standalone benchmarks reward AI for replacing people. Centaur evaluations instead measure what a human–AI team achieves—the uplift over either alone—and thereby make human-centred properties (interpretability, helpfulness, complementarity) first-class objectives rather than afterthoughts. The paper argues that such evaluations are both more demanding and more realistic than current leaderboards, and sets out a research program: who the humans are, how to randomize access, how to score teams. It is the conceptual basis for AlignmentBench (interview-then-predict), SimulacraBench, and the AI-oral-assessment study.
Convex Markov Games: A New Frontier for Multi-Agent Reinforcement Learning
I. Gemp, A. Haupt, L. Marris, S. Liu, G. Piliouras
International Conference on Machine Learning, 2025.
Preferences over occupancy measures: linear in a Markov game, any convex function in a convex Markov game.
Multi-agent reinforcement learning assumes rewards add up over time. Many objectives do not: diversity, imitation, fairness, safety. Convex Markov games let players hold general convex preferences over occupancy measures, strictly generalizing Markov games (including infinite horizon). Pure-strategy Nash equilibria exist, and we approximate them by gradient descent on an exploitability bound. In the prisoner's dilemma, "transient imitation" recovers a profile close to human play with higher utility and roughly a thousand-fold lower exploitability; experiments also show fairer coordination and safer robot-warehouse behaviour.
The Self-Preferencing Prohibition as a Fairness Provision
A. Haupt
ACM Symposium on Computer Science and Law (Poster), 2025.
Ranges across product categories in a third-party seller's records, 2016–2022.
How does a hybrid platform's placement rule—the Buy Box—shape prices, and is Amazon biased toward its own offers? Theory: the consumer-optimal scoring rule is linear in consumer-preference coefficients minus price, and the Reimers–Waldfogel bias test remains valid when prices respond to the rule. Data: a high-volume third-party seller's records for 1,000+ products over 2016–2022. Amazon wins the Buy Box 39–85% of the time by category when it has an offer; the Buy Box is the cheapest offer only 54–63% of the time; an objectively worse offer wins 14–24% of the time.
Risk Preferences of Learning Algorithms
A. Haupt‡, A. Narayanan‡
Games and Economic Behavior 148, p. 415-426, 2024.
Two actions with the same expected payoff: ε-greedy learning ends up choosing the low-variance one.
Standard bandit algorithms are not risk-neutral. ε-greedy learning is emergently risk-averse: between two actions with equal expected payoff it picks the lower-variance one with probability approaching one, and transiently even when the riskier action is strictly better on average. In a recommender or marketplace this favours measurable, homogeneous options and produces inequity and homogenization. Two corrections—reweighting data by how often an action was chosen, or optimistic estimates for rarely-chosen actions—restore risk neutrality.
Formal Contracts Mitigate Social Dilemmas in Multi-Agent Reinforcement Learning
A. Haupt‡, P. Christoffersen‡, M. Damani, D. Hadfield-Menell
Autonomous Agents and Multi-Agent Systems 38, Article 51, 2024.
A contracting stage before the game: agents propose and accept reward transfers, then play.
Add a contracting stage to a Markov game—agents can voluntarily bind themselves to reward transfers—and, with a rich enough contract space, every subgame-perfect equilibrium of a fully observable Markov game is socially optimal; richer spaces weakly raise welfare in general, including under partial observability. The MOCA algorithm makes this learnable, validated in traffic, pollution, and common-pool-resource environments.
Steering No-Regret Learners to a Desired Equilibrium
B.H. Zhang, G. Farina, I. Anagnostides, F. Cacciamani, S.M. McAleer, A. Haupt, A. Celli, N. Gatti, V. Conitzer, T. Sandholm
ACM Conference on Economics and Computation, 2024.
The certifier partitions unobservable quality into priced certificate tiers; sellers choose both their quality and their certificate (schematic).
Certification is the market's answer to unobservable quality, from carbon credits to educational tests to AI models. We study a certifier that designs and prices certificates while sellers choose both quality and certificate and prices form competitively. Under single crossing, sellers' certification choices are uniquely pinned down; the revenue-maximizing certifier's problem reduces to monopoly pricing with non-linear valuations, for which we give an FPTAS; and both welfare- and revenue-optimal certificate menus can be arbitrarily large.
Black-Box Access is Insufficient for Rigorous AI Audits
S. Casper, C. Ezell, C. Siegmann, N. Kolt, T.L. Curtis, B. Bucknall, A. Haupt, K. Wei, J. Scheurer, M. Hobbhahn, et al.
ACM Conference on Fairness, Accountability, and Transparency, 2024.
Human-Centered AI Fellow Stanford Impact Labs Fellow Departments of Computer Science and Economics HAI, DEL, StatsML, CRFM, STAIR Co-Chair, Stanford HAI Affinity Group on Global AI Governance Rising Star in Data Science, 2025 Young Scientist, 8th Economics Lindau Nobel Laureate Meeting, 2025 Young Researcher, 12th Heidelberg Laureate Forum, 2025
Ph.D. Student Engineering-Economic Systems, Computer Science and AI Lab Committee: A. Bonatti, D. Hadfield-Menell, E. Maskin, D. Parkes Co-Chair, MIT AI Ethics and Policy Group Vice-President, MIT Science Policy Initiative Price Theory Summer Camp, Becker Friedman Institute, 2024