inference-time search

Inference-time search means generating several candidate answers and then picking, rather than trusting the first attempt. Because sampling is partly random, the same question asked a few times yields different chains of reasoning; some land on the right answer even when most do not. The simplest form is best-of-N: produce N solutions and keep the best one, judged by majority vote, by a scoring model, or by an exact check when the task allows it, such as running the generated code.

More elaborate forms search the space of reasoning paths the way a chess engine searches moves: branch at promising steps, prune dead ends, and grow only the lines that look strong, sometimes guided by a value model that scores partial progress. The payoff is that many shallow tries can collectively crack a problem no single attempt could. The price is multiplied compute, and a subtle trap: if the way you pick the winner is unreliable, search can confidently surface a wrong answer that happened to be popular.