arXiv cs.CL· Jungseob Lee, Seungyoon Lee, Hyeonseok Moon, Seongtae Hong, Suhyune Son, Sugyeong Eo, Minhyuk Kim, Chanjun Park, Jaehyung Seo·· 13 小时前AI 评分31
Accept or Repair：在对比解码中保留段落支持的答案
Accept or Repair: Preserving Passage-Supported Answers in Contrastive Decoding
AI 导读
研究者提出免训练的 Accept-or-Repair（AoR）方法：若段落支持采样答案则保留，否则用对比解码最多替换两次，并在替换结果与无段落答案重复而段落倾向其他答案时拒绝该替换。实验显示 AoR 在反事实段落上比采样高 12.5 分、普通段落高 2.3 分，在两个场景均超过八个对比解码与提示基线。仅替换段落不支持的初始答案即可提升每种解码器的平均分。
正文 · 原文
Abstract:Large language models often answer from knowledge learned during training, even when a provided passage contradicts that knowledge. To address this problem, contrastive decoding contrasts the predictions under the full and a modified prompt at every token, but it also alters correct answers that the passage supports. We show that checking answers against the passage removes this trade-off. Based on this observation, we propose Accept-or-Repair (AoR), a training-free procedure. AoR keeps a sampled answer if the passage supports it. Otherwise, it tries up to two replacements with contrastive decoding, returning one as soon as it passes this check. It rejects replacements that repeat the answer generated without the passage when the passage favors another answer. Our experiments show that AoR raises the score over sampling by 12.5 points on counterfactual passages and 2.3 on ordinary ones, surpassing eight contrastive and prompting baselines on both, including when each receives the same rejection rule. Replacing only initial answers that the passage does not support improves the average score of every decoder. These results suggest that models can follow passages that contradict their knowledge while preserving accuracy on passages that agree with it. Code will be released at this https URL.
| Comments: | 21 pages, 5 figures. Code will be released at this https URL |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.29727 [cs.CL] |
| (or arXiv:2610.29727v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.29727 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Seungyoon Lee [view email]
[v1]
Wed, 21 Oct 2026 13:32:29 UTC (126 KB)
来源：arXiv cs.CL · arxiv.org