Decision models that output probabilities for typed questions, such as yes/no or multiple-choice, typically require retraining to improve. A new arXiv preprint (2610.25474) proposes a training-free alternative: apply test-time search to a frozen model. The authors show the model a complete proposed answer and then search over possible continuations or choices, scaling computation at inference time.
According to the abstract, this test-time scaling yields a 15× improvement in the probability assigned to the correct "yes" response. The method requires no parameter updates, making it applicable to existing models. The paper is a preprint and has not yet been peer-reviewed; the abstract is truncated, so full experimental details are not available from the source.