提出新方法缓解大模型编辑中的过拟合问题，让知识更新更精准可控。
REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing
- 分两阶段：先提取事实表示，再用方向向量控制隐藏状态调整
- 在多个评测集上显著降低过拟合，同时保持编辑的可靠性与泛化性
- 适合需要精确控制知识更新的场景，如医疗、法律等高精度领域
大语言模型的知识编辑常因过拟合导致事实更新扩散到不相关语境，过度强调编辑内容。为此，我们提出统一的两阶段框架REACT（Representation Extraction And Controllable Tuning），以实现精确且可控的知识编辑。第一阶段通过定制刺激提取潜在事实表示，并结合主成分分析与可学习线性变换，计算每个实例的方向性“信念偏移”向量。第二阶段使用该向量对隐藏状态施加可控扰动，其幅度由预训练分类器门控，仅在上下文需要时才执行编辑。在EVOKE基准上的实验表明，REACT在几乎所有评估指标上显著减少过拟合；在COUNTERFACT和MQuAKE上的测试显示，该方法在多种编辑场景下仍能保持良好的编辑性能（可靠性、局部性和泛化性）。
原文摘要 · Abstract (English)
Large language model editing methods frequently suffer from overfitting, wherein factual updates can propagate beyond their intended scope, overemphasizing the edited target even when it's contextually inappropriate. To address this challenge, we introduce REACT (Representation Extraction And Controllable Tuning), a unified two-phase framework designed for precise and controllable knowledge editing. In the initial phase, we utilize tailored stimuli to extract latent factual representations and apply Principal Component Analysis with a simple learnbale linear transformation to compute a directional "belief shift" vector for each instance. In the second phase, we apply controllable perturbations to hidden states using the obtained vector with a magnitude scalar, gated by a pre-trained classifier that permits edits only when contextually necessary. Relevant experiments on EVOKE benchmarks demonstrate that REACT significantly reduces overfitting across nearly all evaluation metrics, and experiments on COUNTERFACT and MQuAKE shows that our method preserves balanced basic editing performance (reliability, locality, and generality) under diverse editing scenarios.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。