Eliminate training while retaining reliability gains

Determine whether the improved reliability frontier achieved by GracefulRL can be obtained without additional training relative to a standard reinforcement-learning-trained language model.

Background

The paper reports that increasing GracefulRL training compute moves the AIME reliability frontier upward, yielding higher success at a given error rate. This observation raises the unresolved question of whether comparable reliability gains can be achieved without further training, for example through an alternative inference-time procedure or other mechanism. The paper does not resolve this question.