arXiv cs.CL· Yuexin Wu, Dayou Yu, Vasile Rus·· 12 小时前AI 评分18
基于扰动的数据选择方法用于医疗问答
Perturbation-Based Data Selection for Medical Question Answering
AI 导读
研究提出基于扰动的数据选择方法 RSP 与 RSP-Drop，利用高斯扰动最终块表示产生的预测变化频率和非翻转边际收缩两个信号筛选训练子集。在 128 样本预算下，MedGemma-4B 在五个医疗 QA 数据集上的平均 Macro-F1 分别为 47.55 和 48.28，高于熵池匹配随机选择的 45.11。消融显示熵预过滤并非普遍有益，结果仅基于单一骨干模型且未验证临床有效性。
正文
Abstract:Selecting a compact training subset is an important step in adapting medical language models from existing question-answering banks. We investigate perturbation-based data selection using two signals derived from Gaussian perturbations of the final-block representation: prediction-change frequency and non-flipping margin contraction. After a shared entropy prefilter, Representation-Sensitivity Prioritization (RSP) combines the signals using automatic root-mean-square scaling, while RSP-Drop ranks candidates by contraction alone. Both selectors use only questions and answer options during selection. We evaluate them with MedGemma-4B on five medical QA datasets at retained-set budgets of 64, 128, and 256 examples. At the primary budget of 128, RSP and RSP-Drop obtain dataset-averaged Macro-F1 scores of 47.55 and 48.28, respectively, compared with 45.11 for entropy-pool-matched random selection. Their relative performance varies across datasets and budgets. A selected-set case study identifies contrasting clean-margin profiles among the examples exchanged by the two rules, while a full-pool ablation shows that entropy prefiltering is not uniformly beneficial. These retrospective results characterize the empirical utility and limitations of perturbation-based selection under a fixed adaptation protocol. Evaluation uses one backbone and benchmarks with documented training-family overlap; it does not establish clinical effectiveness.
| Comments: | 10 pages. Preprint submitted to IEEE ICHI 2027. Extends the non-archival GenAI4Health @ NeurIPS 2026 study arXiv:2609.09684v2 with automatic RMS scaling, RSP-Drop, budget analyses, and entropy-prefilter ablations |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.33440 [cs.CL] |
| (or arXiv:2610.33440v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.33440 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yuexin Wu [view email]
[v1]
Sat, 24 Oct 2026 04:33:48 UTC (103 KB)
来源：arXiv cs.CL · arxiv.org