PASTA: A Paraphrasing And Self-Training Approach for Knowledge Updating in LLMs

2026-06-27

AI Summary

This work addresses the challenge of efficiently updating specific factual knowledge—such as recent news events—in large language models after pretraining, a task hindered by the models’ tendency to generate hallucinations. To this end, the authors propose PASTA, a novel framework that synergistically integrates data augmentation, question-answer pair generation, and a self-learning Direct Preference Optimization (DPO) mechanism. Through coordinated rewriting and self-training strategies, PASTA jointly optimizes factual knowledge injection while suppressing hallucinatory outputs. Experimental results demonstrate a dramatic improvement in accuracy on post-cutoff-date news question-answering tasks, rising from 0.02 to 0.82, without compromising general language capabilities. The study also provides a systematic analysis of key configuration factors influencing knowledge updating, offering an effective pathway toward dynamic knowledge integration in pretrained models.

Abstract

Knowledge updating in pre-trained Large Language Models (LLMs) remains an important challenge. While continual training provides a potential avenue for knowledge updating, it continues to present substantial technical difficulties. Furthermore, LLMs often struggle with accurately answering questions about specific factual information, such as news articles - a capability limitation widely recognized in the research community. This paper proposes PASTA, a simple yet powerful framework for integrating detailed factual information from news articles as new knowledge into LLMs, with the primary goal of building specialized models that accurately answer questions about this knowledge. Our framework combines data augmentation, question-answering generation, and a novel self-learning DPO process that simultaneously enables knowledge overwriting and hallucination suppression. We provide insights into effective knowledge updating through systematic analysis of learning parameters and data configurations. In our experimental evaluation with web articles published after the base model's knowledge cutoff, PASTA achieved remarkable improvement from 0.02 to 0.82 accuracy while maintaining general language capabilities, demonstrating its effectiveness for creating domain-specialized LLMs.

Problem

• knowledge updating
• Large Language Models
• factual information
• news articles
• hallucination

Innovation

• knowledge updating
• self-training
• DPO
• data augmentation
• hallucination suppression