A-Evolve-Training: Autonomous Post-Training of a 30B Model

The article announces A-Evolve-Training, an autonomous system capable of post-training a 30 billion parameter model, Nemotron, without human intervention. This system successfully completed four rounds of training, achieving a held-out score of 0.86 on the NVIDIA Nemotron-Reasoning Challenge leaderboard, closely competing with human submissions. Notably, it adapted its search policy in response to misleading development metrics, demonstrating the potential for autonomous systems to not only optimize but also discover new training strategies, marking a significant advancement in autonomous machine learning capabilities at scale.