Welfare, Improvability, and Variance: A Principal-Agent Approach to Optimal Benchmark Item Aggregation
This work addresses the under-examined problem of optimal item aggregation in AI benchmarks, which is crucial for researchers and practitioners to create more meaningful and less misleading evaluations.
This paper models AI benchmarking as a multitask principal-agent game, demonstrating that benchmark welfare loss is determined by item-level alignment with normative welfare priorities, marginal improvability, and performance variance. Applying this framework to OLMES items, the authors identify Pareto-inferior items when considering a pro-worker welfare operationalization.
AI benchmarks have well-documented limitations, with prior work examining contamination, saturation, and construct underspecification. Aggregation has received far less attention: benchmarks are typically summarized by uniformly averaging item-level scores, implicitly treating every test item as equally valuable. We model benchmarking as a multitask principal-agent game and show that the welfare loss from a benchmark is determined jointly by three item-level primitives: alignment with normative welfare priorities, marginal improvability, and performance variance. We translate the theory into an audit framework that ranks items along each of these three axes, and apply it to OLMES items using WORKBank for welfare, the EvoLM 4B suite for improvability, and the PolyPythias 410M panel for variance. The framework surfaces items that are Pareto-inferior within OLMES subject to a pro-worker welfare operationalization. All code is available at https://github.com/stair-lab/principal-agent-benchmarks.