arXiv cs.CL· Yongtong Li, Yixuan Florence Wu, Jiawei Zhang, Naichen Shi·· 12 小时前AI 评分29
跨域表征签名自适应迁移，提升 LLM 可靠性检测
Adaptive Transfer of Representational Signatures for LLM Reliability Detection
AI 导读
研究者提出一种跨域迁移表征签名的统计框架，先为每个领域提取低维签名，再用交叉拟合学习各领域预测器的组合权重，权重随源域与目标域的相关性自适应调整。在新领域仅有 25 条标注问题时，该方法将幻觉检测 AUROC 提升 6.8%，推理步骤预测 AUROC 提升 8.4%。实验覆盖多个幻觉数据集与开放权重 LLM。
正文
Abstract:Large language models (LLMs) are reported to be susceptible to reliability issues including hallucinations and reasoning errors. Existing quality detectors can predict unreliable outputs based on internal representations with decent accuracies. However, training these detectors often requires substantial domain-specific labeled data, which limits their applicability to new domains. We propose to address this challenge by transferring representational signatures across related domains. More specifically, we build a statistical framework that first extracts low-dimensional signatures for each domain, then uses cross-fitting to learn composition weights of domain-specific predictors. The weights adapt to each source domain's predictive relevance to the target domain. Our asymptotic analysis of a least-squares analogue characterizes the limiting prediction risk and aggregation weights in terms of domain discrepancies and representation bias. Experiments across multiple hallucination datasets and open-weight LLMs show that, with only 25 labeled questions in the new domain, adaptive transfer improves hallucination-detection AUROC by 6.8%. For reasoning-step prediction, it improves AUROC by 8.4%.
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.33442 [cs.CL] |
| (or arXiv:2610.33442v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.33442 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yongtong Li [view email]
[v1]
Sat, 24 Oct 2026 04:34:46 UTC (799 KB)
来源：arXiv cs.CL · arxiv.org