arXiv cs.CL· Phongsakon Mark Konrad, Toygar Tanyel, Serkan Ayvaz·· 11 小时前AI 评分34
Routing Subspaces：审计微调语言模型中的评测与部署行为错配
Routing Subspaces: Auditing Evaluation-to-Deployment Mismatch in Fine-Tuned Language Models
AI 导读
研究者提出 Routing Subspaces 审计方法，在路径修补定位的中层窗口拟合配对激活对比，再修改该坐标以缩小微调模型的评测与部署行为差距。在四个全矩阵指令微调模型实例的十二个模型—行为设定中，十个设定的差距缩小超过 0.12，四个设定缩小至 0.06 以下。该方法为非自适应白盒诊断工具，不构成训练时防御或部署安全保证。
正文 · 原文
Abstract:Safety evaluations often assume that behavior observed during testing reflects behavior in ordinary use, but fine-tuning can break this assumption. A checkpoint can appear fixed under evaluation-style prompts while the same behavior persists under ordinary-use prompts. Output scores reveal this mismatch but do not locate it. We investigate whether the distinction is encoded in a stable internal site and introduce an audit that fits a paired activation contrast at a path-patching-informed mid-depth window, then modifies the resulting coordinate on held-out prompts. The intervention reduces the evaluation-to-deployment gap by more than $0.12$ in ten of twelve model--behavior settings and closes it (post-intervention gap below $0.06$) in four, across four full-matrix instruction-tuned model instances; among the eight settings with $n{\geq}120$ paired questions, six reduce and one closes. A fifth model supports localization and edit-provenance checks, and in the twelve settings deployment-framed rates change by at most $6.1$pp. On retrained checkpoints, random directions at the same dose do not reproduce the effect, and off-window controls do so only in part. The strongest literal evidence for an evaluation-versus-deployment distinction comes from sandbagging, whose prompt framing is explicitly evaluative; sycophancy and refusal test framing-contingent task distinctions. The two weak cells, both sycophancy, show where a single-coordinate audit at a fixed window is not sufficient. The audit is a non-adaptive, white-box diagnostic for controlled fine-tuned checkpoints, not a training-time defense, a guarantee of deployment safety, or evidence about naturally arising failures.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Software Engineering (cs.SE) |
| Cite as: | arXiv:2607.20436 [cs.CL] |
| (or arXiv:2607.20436v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2607.20436 arXiv-issued DOI via DataCite |
Submission history
From: Phongsakon Mark Konrad [view email]
[v1]
Mon, 11 May 2026 13:44:35 UTC (426 KB)
[v2]
Sun, 1 Nov 2026 11:27:00 UTC (285 KB)
来源：arXiv cs.CL · arxiv.org