Sampling multiple solutions spends computation on intermediate deductions and unfinished arguments, leaving the final answers under‑utilized. To address this, we introduce ReSolve, a training‑free inference procedure that reuses existing candidate reasoning through selective generative moderation. The core component is an answer‑distribution controller: when candidates disagree or lack a parseable answer, the controller asks the model to examine the current derivations and injects the generated solution into a bounded loop.
We evaluate on 130 competition‑mathematics problems using Hybrid scoring with two independently sampled candidate pools. ReSolve achieves 100 correct answers in the first pool and 99 in the second, whereas voting over the same four candidates yields 91 and 92 correct answers respectively, with no correct‑to‑incorrect flips.
Eight‑sample self‑consistency obtains 94 and 96 correct answers but consumes substantially more tokens; ReSolve reduces token usage by 46.3% and 47.2% in the two evaluations.
A controlled ablation that removes visible derivations while keeping answer keys, vote counts, and the per‑state output‑cap rule drops accuracy from 100 to 93, even though computation increases.
Both selective and always‑on Uniform moderation solve 97 problems; selectivity cuts moderation tokens by roughly 54% and lowers total pipeline tokens by 6.2% compared with always‑on moderation.
These findings support candidate reasoning as reusable inference computation, but they do not establish an accuracy advantage over additional sampling or a distinct benefit from specialized route instructions.
Review