New research reveals that machine learning research agents avoid overfitting by learning highly compressible strategies that fit into just a few tokens, mirroring the behavior of human research communities. This mechanism prevents rote memorization and explains why benchmark-driven AI continues to generate genuine real-world advancements.