蒸馏小语言模型的弃权选项测量研究：拒绝损失、采样抖动与选项俘获
What the Abstention Option Measures: Refusal Loss, Sampling Churn, and Option Capture in Distilled Small Language Models
AI 导读
一项针对蒸馏小语言模型的测量研究拆解了含弃权选项的偏见基准分数背后的三种机制：拒绝损失、采样抖动和选项俘获。研究发现，参考模型在不同推理种子间重采样时，10% 到 27% 的拒绝会变成刻板印象答案，未计此基线的逐项拒绝指标会夸大损害；而用拒绝样本训练的学生模型在弃权从不正确的科学题上最多 95% 会选弃权选项，并迁移到三个额外基准。
正文 · 原文
Abstract:Small instruction-tuned language models reach phones, cars, and offline assistants through knowledge distillation, and whether a distilled model is fair enough to ship is often decided by a multiple-choice bias score on a benchmark that offers an abstention option. Three mechanisms move behavior around the abstention option: refusal loss, where the student stops declining on items its pre-distillation reference declined; sampling churn, the same change observed between two runs of the reference itself; and option capture, where the student takes an abstention option whenever one is offered. Injecting each mechanism into a reference's own records shows that aggregate scores see only outright loss, per-item scores see churn and remapping, and only conditioning on committed answers stops moved refusal mass from reading as fairness. Resampling a reference across inference seeds already turns 10% to 27% of its refusals into stereotype answers, so a per-item refusal metric reported without that floor overstates damage. Option capture is what corpus repair teaches: students trained with refusal exemplars take an offered abstention option on up to 95% of science questions where abstention is never correct, and the habit carries to three further benchmarks. Free-form evaluation shows that the refusals multiple choice loses are not where free-form stereotyping occurs. We distill these findings into a per-condition report anchored to the reference, with measured floors and ceilings, and apply it to four student families against a fine-tuning control: distillation keeps fewer of the reference's refusals than fine-tuning on the same prompts where the reference's refusals are reliable, by a few points, while the common metrics report improvements whenever refusal mass moves. A bias score quoted for a distilled model should say which of these it reports.
| Comments: | 29 pages, 9 figures, 15 tables. v3: retitled (previously "The Asymmetric Effects of Knowledge Distillation on Bias in Small Language Models") and reframed as a measurement study |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY) |
| ACM classes: | I.2.7; I.2.6; K.4.1 |
| Cite as: | arXiv:2607.28639 [cs.CL] |
| (or arXiv:2607.28639v3 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2607.28639 arXiv-issued DOI via DataCite |
Submission history
From: Plawan Kumar Rath [view email]
[v1]
Tue, 19 May 2026 23:09:28 UTC (88 KB)
[v2]
Mon, 3 Aug 2026 22:51:27 UTC (96 KB)
[v3]
Sat, 24 Oct 2026 15:37:00 UTC (147 KB)
来源：arXiv cs.AI · arxiv.org