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

TL;DR

REACT is a two-phase framework that extracts factual representations and applies controllable perturbations to improve knowledge editing in large language models, significantly reducing overfitting and maintaining balanced editing performance.

Contribution

It introduces a novel method combining representation extraction and controllable tuning to address overfitting in LLM knowledge editing.

Findings

REACT significantly reduces overfitting across multiple benchmarks.

The method preserves balanced editing performance under diverse scenarios.

Experiments demonstrate improved reliability, locality, and generality.

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.