Reproduce “KnowledGPT: Enhancing Large Language Models with Retrieval and Storage Access on Knowledge Bases” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xintao Wang, Qianwen Yang, Yongting Qiu, Jiaqing Liang, Qianyu He, Zhouhong Gu, Yanghua Xiao, Wei Wang

KnowledGPT, developed at Fudan University, introduces a framework enabling Large Language Models to interact with structured Knowledge Bases for both retrieval and storage. The system generates Python code for programmatic KB access and manages personalized knowledge, achieving F1 scores of 0.92 on single-hop and 0.93 on multi-hop KBQA datasets while increasing knowledge coverage through expanded representations.

Key insights

  • Program of Thought (PoT) prompting allows LLMs to generate executable code for programmatic interaction with KBs, enabling complex, multi-step reasoning over structured data.
  • Robust LLM-based entity linking, aided by candidate entity descriptions, resolves ambiguities and improves the accuracy of KB interactions.
  • Personalized Knowledge Bases (PKBs) with entity-aspect information significantly expand knowledge coverage (from 53% to over 80% word recall) and enable adaptable, writable symbolic memory for LLMs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.