Repeating an instruction once boosts how reliably AI follows it.

Language models often fail to follow multi-step rules, especially when other systems need to inspect or repair their work. To fix this, researchers tested a simple trick called instruction duplication. Instead of retraining the model or changing how it picks words, they simply repeated the procedural instruction twice during testing. This gives the model a second chance to catch the rule. Repeating the instruction raised success on an eight-test diagnostic from 90.22% to 93.17%, removing 30.2% of the errors left after a single copy. Vocabulary recall also improved from 73.44% to 74.81%. Final-answer accuracy stayed exactly 60.21%, but jumping to conclusions early grew from 1.52% to 2.30%. The boost matters most when another program uses the model's intermediate steps. In one repair task, success jumped from 84.2% to 97.1%. Repeating an instruction is an easy control that helps downstream tools act on generated paths.

Why it matters

This matters for developers and users of AI systems that need precise step-by-step execution, like in medical diagnostics or automated repairs. It reduces errors when outputs must be checked or fixed later, making control simpler.

Questions answered