Making the model show its work.
LLMs can solve harder problems when they generate intermediate steps instead of jumping to the answer. Adding “let’s think step by step” to a prompt is enough to trigger this. The model breaks a multi-step problem into pieces, solves each one, and arrives at a final answer. In early experiments, chain-of-thought prompting roughly tripled accuracy on math and logic benchmarks for large models.
Originally described as an emergent ability that only appeared above ~100B parameters. Smaller models have since shown some reasoning capability, especially with better training data.
Models like OpenAI’s o1 internalize this pattern through RL training rather than relying on the prompt. The chain of thought happens inside the model’s generation, not because you asked for it.