Verify rather than think

There is often a more profitable path than extending reasoning time, and it rests on a fundamental asymmetry.

The asymmetry

Verifying is easier than producing. Finding the exit of a maze is hard; checking that a proposed path is valid is immediate. This asymmetry holds for a large share of useful problems.

Generate several times, then decide

Produce five independent responses at non-zero temperature, then keep the one that recurs most often. On single-answer problems, the gain is clear and well documented: errors are scattered, correct answers converge.

Cost: five times the inference, but with a fast model. That is often cheaper than a single call to a reasoning model, and sometimes better.

A separate verifier

Even better when possible: a second call whose only job is to check the first one's proposal, with the explicit instruction to hunt for the error. A model asked to "verify" approves; one asked to "find what's wrong" finds it.

And when a deterministic check exists, running the code, recalculating the total, validating the format, use it. It is free, instant, and it does not make mistakes.

The limit

On a question with no right answer, a judgment, a style, a strategy, majority voting means nothing. It selects the most consensual, i.e. the most bland. These methods are for problems that have a solution, not for those that have an opinion.