Asking an AI to reason through intermediate steps can improve its answers, but the benefit depends on what those steps contain. A study offers an explanation for both the gains and the failures.
Published in July in the Journal of Machine Learning Research, the paper analyses chain-of-thought prompting through a statistical model of multistep tasks. Under its assumptions, the method behaves like Bayesian inference, using examples in a prompt to infer the task before producing an answer.
That framing connects directly with work examining how hidden reasoning steps operate inside leading AI models, where intermediate reasoning traces can reveal more about model behavior than the final answer alone. It also contrasts with approaches that move reasoning away from explicit text, such as systems that perform more of their reasoning in an internal latent state.
The researchers separate errors into two sources: limitations inherited from pretraining and uncertainty introduced when the model interprets the prompt. Their analysis shows that the latter can decrease exponentially as the number of examples grows, under its assumptions.
There is a qualification. Intermediate steps that carry too little relevant information can make chain of thought perform worse than ordinary prompting. Experiments support the distinction between helpful reasoning examples and distracting or incomplete ones. That distinction matters for the wider challenge of building AI systems whose reasoning is reliable rather than merely fluent.
The findings do not guarantee that longer answers are better, or that models follow the theory exactly. They offer a practical test for prompt design: do the added steps actually help identify the solution?