In 1927, psychologist L. L. Thurstone introduced a concept in his paper, “A law of comparative judgment,” suggesting that individuals choose the option with the highest value to them among several alternatives, despite not being able to assign a specific value to their choice. Thurstone’s work laid the foundation for psychometrics, which aims to measure mental processes, and for what are now known as random utility models (RUMs). These models offer a mathematical approach to understanding human preferences, which can help predict outcomes in hypothetical scenarios.<\/p>\n
RUMs evaluate the “utility” or benefit of a choice, such as selecting which book to read first from a library stack. Gabriele Farina, an assistant professor at MIT, notes that these models are inherently random because individual preferences can vary. For instance, someone might typically prefer coffee over tea in the morning but switch preferences occasionally. RUMs are widely used in government and industry to predict behaviors in scenarios like transportation choices if a road is closed or how to allocate funds from a financial windfall.<\/p>\n
Though RUMs have been in development for nearly a century, a recent paper presented at the International Conference on Learning Representations in Brazil revealed new insights. The paper, authored by researchers including MIT’s Yeshwanth Cherapanamjeri, Gabriele Farina, Constantinos Daskalakis, and Sobhan Mohammadpour, identified a flaw in how RUMs are typically estimated using pairwise comparisons. This method, which compares two items at a time, fails to reveal correlations between preferences, such as a voter who supports both gun control and government-sponsored child care.<\/p>\n
The MIT team demonstrated that two-way comparisons alone cannot uncover these correlations, but analyzing preferences among three options can. By asking participants to rank three items, they developed a method to integrate these individual rankings into a comprehensive model. According to Farina, their research focuses on creating algorithms to extract preference information efficiently without needing exponentially more data as the number of items increases.<\/p>\n
Emma Frejinger from the University of Montreal praised the study as a breakthrough, showing that collecting best-of-three choices can enhance the accuracy of these models. Constantinos Daskalakis emphasized the ongoing importance of building utility models, which have been vital to the internet economy and will continue to be crucial for aligning AI models. He highlighted that RUMs are essential for the commercial success and functionality of large language models, which rely on understanding user preferences during training. Daskalakis pointed out that as options multiply in various domains, developing models to predict preferences is necessary, with a continuous need for refinement to make accurate predictions.<\/p>\n