A new study explores how language models stabilize their answer preferences before verbalizing responses, revealing that this stabilization occurs earlier than the answer becomes clear. The findings could enhance understanding of model behavior and improve the design of AI systems for reasoning tasks.