This research uncovers three distinct phase transitions in large language model training that mirror human neurological development.

• Models demonstrate hierarchical learning patterns similar to human brain development
• Researchers identified triple phase transitions corresponding to different cognitive abilities
• Internal network states show consistent transition patterns across model scales
• These transitions predict when new capabilities will emerge during training

Medical relevance: Understanding these transitions offers insights into both artificial intelligence and human neurological development, potentially informing neuroscience research and medical applications for cognitive disorders.

Triple Phase Transitions: Understanding the Learning Dynamics of Large Language Models from a Neuroscience Perspective