For tasks involving math or multi-step logic, what technique tends to improve accuracy?

Multi-step problems fail in a particular way: the model commits to an answer in one jump, and a single bad step in the middle is invisible because no middle was ever shown. Asking for the reasoning before the answer — chain-of-thought prompting — makes it work through the steps in order, and each step it writes down conditions the next one.

It also changes what you can do when the answer is wrong. A bare number gives you nothing to check. A worked derivation shows you which line went wrong, and that line is often a misreading of the question rather than the arithmetic.

One thing has changed and is worth knowing. Claude's recent models think adaptively, reasoning before they answer on a hard question whether or not you ask. Asking explicitly still matters when thinking is off, when the model would otherwise answer directly, and when you want the reasoning visible rather than internal. What has not changed is the failure mode of the other three options: less reasoning and less context, on exactly the tasks that need both.