Robust AI Evaluation through Maximal Lotteries

A team of researchers from Harvard, MIT, and Meta has published a paper titled 'Robust AI Evaluation through Maximal Lotteries,' challenging the standard practice of ranking AI models. The current method, which aggregates human preferences via pairwise comparisons into a single Bradley-Terry ranking, is fundamentally flawed. It forces heterogeneous human judgments into a simplistic total order, violating core social-choice principles and failing to represent the diversity of what users consider 'better.' The paper argues this pushes the field toward a monoculture of models optimized for a narrow, averaged preference.

The researchers demonstrate that existing social-choice alternatives, like maximal lotteries, are too sensitive to data variation and can promote models that severely underperform for specific subpopulations. Their solution is 'robust lotteries,' a novel aggregation method that optimizes for worst-case performance under plausible shifts in preference data. Tested on large-scale datasets, robust lotteries provide more reliable win-rate guarantees across the full annotator distribution and recover a stable, pluralistic set of top models. This shift from a single ranking to a set of complementary winners is a principled step toward an AI ecosystem that serves the full spectrum of human values and tasks, moving beyond the misleading chase for a single 'best' model.