arXiv：cs.AI· Bangji Yang, Jiajun Fan, Hongbo Ma, Xi Zhu, Weizhi Zhang, Minghao Guo, Ye Li, Hamid Palangi, Jiaxuan You·· 5 小时前AI 评分40
ReSolve：通过选择性生成式审核复用候选推理
ReSolve: Reusing Candidate Reasoning through Selective Generative Moderation
AI 导读
ReSolve 是一种无需训练的推理方法，通过选择性生成式审核复用候选推理过程。在 130 道竞赛数学题上，Hybrid 评分下两轮评测分别答对 100 和 99 题，而同样四个候选的投票仅答对 91 和 92 题，且 token 消耗比 8 样本自一致性减少 46.3% 和 47.2%。
正文
Abstract:Sampling multiple solutions spends computation on intermediate deductions and unfinished arguments as well as final answers. We introduce ReSolve, a training-free inference procedure that reuses this candidate reasoning through selective generative moderation. An answer-distribution controller invokes a model to examine existing derivations when candidates disagree or lack a parseable answer, then incorporates the generated solution into a bounded loop. Under Hybrid scoring on 130 competition-mathematics problems evaluated with two independently sampled candidate pools, ReSolve obtains 100 and 99 correct answers, compared with 91 and 92 for voting over the same four candidates, with no correct-to-incorrect changes relative to that vote in either pool. Eight-sample self-consistency obtains 94 and 96 correct answers while consuming substantially more tokens; ReSolve uses 46.3% and 47.2% fewer tokens in the two evaluations. A controlled ablation removes visible derivations while retaining answer keys, vote counts, and the per-state output-cap rule, reducing accuracy from 100 to 93 correct despite increasing computation. Selective and always-on Uniform moderation both solve 97 problems, while selectivity reduces moderation tokens by approximately 54% and total pipeline tokens by 6.2%. These results support candidate reasoning as reusable inference computation. They do not establish an accuracy advantage over additional sampling or a distinct benefit from specialized route instructions.
| Comments: | Corrected a typo in an author's name. No changes to the paper content |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.01140 [cs.AI] |
| (or arXiv:2610.01140v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.01140 arXiv-issued DOI via DataCite |
Submission history
From: Bangji Yang [view email]
[v1]
Thu, 1 Oct 2026 06:19:29 UTC (193 KB)
[v2]
Tue, 6 Oct 2026 05:27:30 UTC (193 KB)
来源：arXiv：cs.AI · arxiv.org