Red Queen G"odel Machine Co-Evolves AI Agents and Evaluators

The Red Queen G"odel Machine (RQGM) is a novel evolutionary framework designed to address a critical limitation in current AI self-improvement methods: the assumption of static evaluation criteria. Unlike traditional approaches that rely on fixed benchmarks, RQGM allows both AI agents and their evaluators to evolve simultaneously. This dynamic co-evolution mirrors natural selection, where species adapt as their environments change. The framework organizes search into epochs, with evaluation criteria updated at epoch boundaries, ensuring self-improvement guarantees within each epoch while the overall objective evolves. The researchers demonstrated RQGM's effectiveness by improving test pass rates on coding tasks, using fewer tokens than prior state-of-the-art methods. Furthermore, RQGM significantly enhanced performance in complex domains like scientific paper writing and reviewing, and Olympiad-level proof writing. Co-evolved writers achieved higher acceptance rates, and graders showed improved ground-truth accuracy. Notably, the system corrected a bias in baseline reviewers that over-accepted AI-generated papers by introducing an adversarial objective.