Invited Talk: The Art & Science of Benchmarking Agents (Fred Sala)
Abstract
Our ability to measure AI has been outpaced by our ability to develop it, and this eval gap is one of the most important problems in AI. We need more enduring benchmarks to close this gap, and consequently advance entire new vectors of capabilities for the field. In this talk, I'll share our insights into evaluating agents, drawing from experience working with nearly all frontier labs and many of our academic collaborators. We'll discuss the science (i.e., mechanics that make benchmarks rigorous and effective) and art (i.e., intangibles driving ambitious and enduring benchmarks) of building great benchmarks. I'll close by sharing some of the learnings from Open Benchmarks Grants— a $3M initiative in partnership with Hugging Face, Together AI, Prime Intellect, Factory, and others— and highlighting some of the projects we're most excited about funding.