How Do Large Language Models Capture the Ever-changing World Knowledge? A Review of Recent Advances

TL;DR

This survey addresses how to keep large language models aligned with the ever-changing world knowledge without re-training. It provides a taxonomy that divides approaches into implicit methods (knowledge editing and continual learning) and explicit methods (memory-enhanced, retrieval-enhanced, and internet-enhanced augmentation), reviewing representative works in each category. The authors compare methods, discuss challenges such as knowledge conflicts, efficiency, and evaluation, and outline future directions to guide research and practice in dynamic knowledge tasks. By clarifying trade-offs and proposing benchmarks, the paper aims to inform when to apply editing versus retrieval-based strategies and how to evaluate dynamic knowledge updates in deployed LLMs.

Abstract

Although large language models (LLMs) are impressive in solving various tasks, they can quickly be outdated after deployment. Maintaining their up-to-date status is a pressing concern in the current era. This paper provides a comprehensive review of recent advances in aligning LLMs with the ever-changing world knowledge without re-training from scratch. We categorize research works systemically and provide in-depth comparisons and discussion. We also discuss existing challenges and highlight future directions to facilitate research in this field. We release the paper list at https://github.com/hyintell/awesome-refreshing-llms

Figures (6)

• Figure 1: A trained LLM is static and can be outdated (e.g., ChatGPT; OpenAI_chatgpt2022). How can LLMs be aligned to the ever-changing world knowledge efficiently and effectively?
• Figure 2: Taxonomy of methods to align LLMs with the ever-changing world knowledge (due to space limitation, please refer to Appendix \ref{['append_complete_taxonomy']} for a complete review). Implicit means the approaches seek to directly alter the knowledge stored in LLMs (e.g., parameters) (\ref{['implicitly_align']}), while Explicit means more often incorporating external resources to override internal knowledge (e.g., search engine) (\ref{['explicitly_align']}).
• Figure 3: A high-level comparison of different approaches.
• Figure 4: Single-Stage (left) typically retrieves once, while Multi-Stage (right) involves multiple retrievals or revisions to solve complex questions (\ref{['retrieval_enhanced']}).
• Figure 5: An example of knowledge conflict of ChatGPT OpenAI_chatgpt2022. Even if the correct context is provided, ChatGPT still favours its internally memorized knowledge. The screenshot was taken in May 2023 for GPT-3.5 without web browsing.