Reproduce “FacLens: Transferable Probe for Foreseeing Non-Factuality in Fact-Seeking Question Answering of Large Language Models” 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
Yanling Wang, Haoyang Li, Hao Zou, Jing Zhang, Xinlei He, Qi Li, Ke Xu
FacLens introduces a transferable, lightweight probe that foresees non-factuality in Large Language Models' (LLMs) fact-seeking question answering by analyzing their hidden question representations. This approach, which uses unsupervised domain adaptation, achieves competitive non-factuality prediction performance with substantially greater efficiency and transferability across different LLMs compared to existing methods.
Key insights
- Hidden question representations within an LLM contain sufficient information to effectively predict non-factuality (knowledge awareness) before response generation.
- Different LLMs, despite varied architectures, exhibit similar underlying non-factuality prediction patterns in their hidden question representations, which facilitates cross-model transferability.
- A lightweight probe combined with unsupervised domain adaptation offers a highly efficient and scalable alternative to resource-intensive methods like fine-tuning for non-factuality prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.