Content area
Several current preference aggregation challenges — ranging from aligning large language models to summarizing civic discussions — closely resemble classical questions in social choice theory. Yet they often fall outside the scope of traditional frameworks: the outputs may be unconventional, individuals may express preferences over only a tiny fraction of alternatives, or real-world instances may exhibit much richer structure. This thesis bridges that gap by developing new theoretical frameworks, formal analyses, and algorithms tailored to these settings, offering principled approaches to collective decision-making in modern, complex environments.
Part I addresses AI alignment. We introduce an axiomatic framework for evaluating alignment methods and show that widely-used loss-minimization techniques violate some of the most fundamental principles of preference aggregation. We further demonstrate that aligning a single model faces inherent limitations, regardless of the algorithm used. To overcome this challenge, we propose using an ensemble of models that collectively capture a broader range of human preferences and provide both theoretical and empirical evidence for the effectiveness of this approach.
Part II focuses on preference elicitation in settings where individuals can feasibly provide only limited feedback across a large space of alternatives, an increasingly common restriction in contexts such as civic participation platforms. We design algorithms that elicit the right information for meaningful aggregation, with provable guarantees, and develop theoretical tools to characterize the fundamental trade-offs imposed by such informational constraints.
Part III studies liquid democracy, a flexible voting system in which individuals may delegate their votes to others. While much of the literature emphasizes worst-case failures, we propose a semi-random model that better captures realistic delegation behavior. Under mild assumptions, we show that liquid democracy avoids many of the previously-identified pitfalls. These results are further supported by real-world experiments in classroom and corporate settings, offering empirical evidence of the paradigm’s practical viability.