2025 Past Seminars

Host: Marcus Pivato

Abstract. A decision theory is self-undermining if, when you ask it which decision theory you should use, it does not consider itself to be among the rationally permissible ones. We distinguish different versions of this property and show that all of the proposed decision theories for imprecise credences are self-undermining in one sense, while the most discussed risk-sensitive decision theories are self-undermining in a slightly different sense. We argue that, in both cases, this gives some reason to reject these theories, though there are escape routes for their proponents.

(Joint work with Catrin Campbell-Moore and Jason Konek)

Host: Lirong Xia

Abstract. In recent years there has been an explosion in interest in topics that sit at the intersection of applications of computing technology and societal issues. There has been significant work in the academic, industrial, and policy spaces to clarify and formalize best practices regarding the deployment of computational decision making (e.g., artificial intelligence and machine learning) at scale. Part of this work has been a newfound interest in many age old conversations about the roles and limits of technology and society. In this talk I’ll give an overview of recent work covering several projects that use tools from computational social choice, data science, and artificial intelligence more generally to build systems and algorithms that are context aware, flexible, and human-centered. This work includes multi-stakeholder recommender systems that use social choice to balance competing fairness and efficiency objectives and online decision making systems that use (inverse) reinforcement learning and human cognitive models for making choices in complex environments.

Host: Marcus Pivato

Abstract. We initiate a novel direction in randomized social choice by proposing a new definition of agent utility for randomized outcomes. Each agent has a preference over all outcomes and a quantile parameter. Given a lottery over the outcomes, an agent gets utility from a particular representative, defined as the least preferred outcome that can be realized so that the probability that any worse-ranked outcome can be realized is at most the agent’s quantile value. In contrast to other utility models that have been considered in randomized social choice (e.g., stochastic dominance, expected utility), our quantile agent utility compares two lotteries for an agent by just comparing the representatives, as is done for deterministic outcomes. We revisit questions in randomized social choice using the new utility definition. We study the compatibility of efficiency and strategyproofness for randomized voting rules, efficiency and fairness for randomized one-sided matching mechanisms, and efficiency, stability, and strategyproofness for lotteries over two-sided matchings. In contrast to well-known impossibilities in randomized social choice, we show that satisfying the above properties simultaneously can be possible.

Joint work with Sanjukta Roy (Indian Statistical Institute)

Host:  Marcus Pivato

Abstract. We propose a new class of social quasi-orderings in a variable-population setting.  In order to declare one utility distribution at least as good as another, the critical-level utilitarian value of the former must reach or surpass the value of the latter.  For each possible absolute value of the difference between the population sizes of two distributions to be compared, we specify a non-negative threshold level and a threshold inequality.  This inequality indicates whether the corresponding threshold level must be reached or surpassed in the requisite comparison.  All of these threshold critical-level utilitarian quasi-orderings perform same-number comparisons by means of the utilitarian criterion.  In addition to this entire class of quasi-orderings, we axiomatize two important subclasses.  The members of the first subclass are associated with proportional threshold functions, and the well-known critical-band utilitarian quasi-orderings are included in this subclass.  The quasi-orderings in the second subclass employ constant threshold functions; the members of this second class have, to the best of our knowledge, not been examined so far.  Furthermore, we characterize the members of our class that (i) avoid the repugnant conclusion; (ii) avoid the sadistic conclusions; and (iii) respect the mere-addition principle.

 (Joint work with Walter Bossert and Kohei Kamaga)

Host: : Danilo Coelho

Abstract. We define and study obvious strategy-proofness with respect to a partition of the set of agents. It encompasses strategy-proofness as a special case when the partition is the coarsest one and obvious strategy-proofness when the partition is the finest. For any partition, it falls between these two extremes. We establish two general properties of this new notion and apply it to the simple voting problem with two alternatives and strict preferences.

(Joint work with Jordi Massó and Alejandro Neme)

Host: Marcus Pivato

Abstract. This paper aggregates individuals’ preferences over menus of lotteries into social preferences. Specifically, we consider the case where both individuals and society have preferences for flexibility (Dekel et al., 2001). That is, they face uncertainty regarding future states that determine their tastes over lotteries, leading them to prefer larger menus at the current time. Each individual’s future tastes are influenced by different aspects of future states, implying that each has their own state space. This paper axiomatically characterizes how society should construct a collective state space, which represents the entire set of factors influencing society’s future tastes. All of our axioms are motivated by discussions on the Pareto principle.

Host:  Marcus Pivato

Abstract. Consider the problem of aligning large language models (LLMs) with human values. The standard approach begins with pairwise comparisons from users of the form "between these two outputs to the prompt, which do you prefer?" This response data is aggregated into a reward function, giving numerical scores to outputs, which is subsequently used to steer an existing LLM toward higher-reward answers. This process is essential for making LLMs helpful while avoiding dangerous or biased responses.

However, this paradigm faces a fundamental challenge: people often disagree on what constitutes a "better" output. What, then, should we do when faced with diverse and conflicting preferences?

This talk explores two approaches to this challenge rooted in social choice theory. First, we take an axiomatic perspective, arguing that the process of learning reward functions should satisfy minimal requirements such as Pareto Optimality: if all users unanimously prefer one outcome to another, the aggregated reward function should reflect this. We show that current alignment methods necessarily violate these basic axioms. In contrast, we provide a proof-of-concept aggregation rule that is guaranteed to satisfy them. Second, we explore a more radical approach: representing, rather than resolving, disagreement. Instead of training a single LLM, we train an ensemble, analogous to multi-winner voting systems. We introduce a novel criterion, pairwise calibration, inspired by proportionality. Together, these approaches provide a principled foundation for building AI systems aligned with the pluralism of human values. 

(Joint work with Luise Ge, Evi Micha, Ariel D. Procaccia, Itai Shapira, Yevgeniy Vorobeychik, and Junlin Wu.)