Latent Paraphrasing: Perturbation on Layers Improves Knowledge Injection in Language Models

TL;DR

This paper introduces LaPael, a latent-level paraphrasing technique that applies input-dependent noise to LLM layers, improving knowledge injection efficiency and diversity without high computational costs.

Contribution

LaPael is a novel method that performs internal model perturbations for knowledge injection, reducing costs and increasing diversity compared to traditional paraphrasing approaches.

Findings

LaPael outperforms standard fine-tuning on QA benchmarks.

Combining LaPael with data-level paraphrasing yields further improvements.

LaPael reduces paraphrasing costs by eliminating external model usage.

Abstract

As Large Language Models (LLMs) are increasingly deployed in specialized domains with continuously evolving knowledge, the need for timely and precise knowledge injection has become essential. Fine-tuning with paraphrased data is a common approach to enhance knowledge injection, yet it faces two significant challenges: high computational costs due to repetitive external model usage and limited sample diversity. To this end, we introduce LaPael, a latent-level paraphrasing method that applies input-dependent noise to early LLM layers. This approach enables diverse and semantically consistent augmentations directly within the model. Furthermore, it eliminates the recurring costs of paraphrase generation for each knowledge update. Our extensive experiments on question-answering benchmarks demonstrate that LaPael improves knowledge injection over standard fine-tuning and existing noise-based approaches. Additionally, combining LaPael with data-level paraphrasing further enhances performance.