DuoDiscover: Scaling Autonomous Discovery Beyond the Evaluation Bottleneck (arxiv.org)

DuoDiscover, a new search system for autonomous discovery in AI, addresses the significant challenge of evaluation bottlenecks that hinder real-world engineering advancements. Traditional evaluation methods can be highly time-consuming, often taking hours to complete, limiting the number of evaluations that can be feasibly conducted. DuoDiscover innovatively sidesteps this issue by enabling a team of coding agents to jointly determine optimization targets and develop tailored proxy evaluators, referred to as yardsticks. This allows agents to focus on optimizing their progress via quicker yardstick assessments, while fidelity tests continually validate these proxy measures to ensure their effectiveness. The implications for the AI/ML community are substantial, particularly in fields requiring rapid iteration and assessment, such as software and chip design. DuoDiscover has demonstrated impressive results, achieving 1.92x and 1.38x performance improvements on two heavily optimized software systems. On a complex CPU microarchitecture optimization task, it outperformed previous models and human designers, achieving significant gains in much tighter timelines. This advancement not only accelerates the optimization process but extends the capabilities of autonomous systems, paving the way for faster, more efficient discoveries across various applications in AI and machine learning.