arXiv cs.CL· Stephen Peng, Vineet Jagadeesan Nair, Noah Broestl, Anne Friederike Kleppe, Steven Mills, Adi Zolotov·· 11 小时前AI 评分40
生物医学领域微调大语言模型的安全护栏侵蚀：知识型弃答的配对审计
Guardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMs
AI 导读
研究审计了 25 个生物医学领域微调大语言模型与 11 个配对基座检查点，发现知识型适当弃答在微调后进一步下降：MedAbstain 上中位数从基座的 8.3% 降至 4.2%，25 对中有 21 对出现下降。变化幅度与方向在不同模型对之间差异明显，作者据此呼吁在提升领域能力的同时保留适当谨慎的微调方法。
正文 · 原文
Abstract:The number of domain fine-tuned models, particularly in the biomedical domain, is rapidly expanding. Fine-tuning promises high-performing models at a fraction of the cost of training a base model, but it can also produce unintended behavioral changes. Multiple studies have shown that safety guardrails can erode with fine-tuning, yet large-scale studies across model architectures, data mixtures, and fine-tuning methodologies remain limited. Where prior work in this space has studied value-centric refusal, this work focuses on knowledge-centric abstention. We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench to evaluate the impact of fine-tuning on knowledge-centric abstention. We find that defined appropriate abstention measured by these benchmarks, already low in base models (median 8.3% on MedAbstain), falls further after fine-tuning (median 4.2%, decreasing in 21 of 25 pairs on MedAbstain), and that the magnitude and direction of change are heterogeneous across pairs. These results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.
| Comments: | To be published in the Proceedings of the 6th Machine Learning for Health Symposium (ML4H 2026), PMLR volume 346 |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.08780 [cs.CL] |
| (or arXiv:2611.08780v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.08780 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Stephen Peng [view email]
[v1]
Thu, 5 Nov 2026 23:03:11 UTC (619 KB)
来源：arXiv cs.CL · arxiv.org