REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing

2025-05-25

AI Summary

To address pervasive overfitting in large language model (LLM) knowledge editing—where factual updates propagate beyond the target scope and excessively activate in irrelevant contexts—this paper proposes a two-stage controllable editing framework. The method first extracts direction vectors grounded in “belief shift” to precisely identify the knowledge representation subspace for editing. Second, it introduces a context-aware gating perturbation mechanism that jointly modulates hidden-state perturbation magnitude via a learnable linear transformation and a pretrained classifier. Evaluated on standard benchmarks, the approach significantly suppresses overfitting on EVOKE, while achieving high reliability, strong locality, and robust generalization on COUNTERFACT and MQuAKE. It unifies editing accuracy, necessity, and controllability—advancing both fidelity and interpretability in LLM knowledge modification.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

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.

Problem

• Overfitting in large language model knowledge editing methods
• Uncontrolled propagation of factual updates beyond intended scope
• Balancing precision and controllability in knowledge edits

Innovation

• Extracts latent representations using tailored stimuli
• Applies controllable perturbations with belief shift vectors
• Gates edits via pre-trained classifier for contextual necessity