Reproduce “KGLens: Towards Efficient and Effective Knowledge Probing of Large Language Models with Knowledge Graphs” 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
Shangshang Zheng, He Bai, Yizhe Zhang, Yi Su, Xiaochuan Niu, Navdeep Jaitly
The KGLens framework from Apple introduces an efficient and effective method for evaluating Large Language Models' factual knowledge by leveraging knowledge graphs. It significantly reduces the computational resources needed for probing by 35% compared to brute-force methods while generating highly sensible questions (97.7%) and verifying answers with 96% human agreement, enabling granular identification of knowledge blind spots.
Key insights
- Thompson sampling, particularly when combined with signal propagation, substantially reduces the computational cost of identifying an LLM's knowledge deficiencies.
- Leveraging GPT-4 within a graph-guided question generation strategy produces highly sensible and unambiguous queries, validated by human evaluation.
- Fine-grained analysis using KGLens's metrics and visualizations can reveal specific knowledge blind spots, such as temporal biases or differential performance across entity groups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.