Reproduce “HalluEntity: Benchmarking and Understanding Entity-Level Hallucination Detection” 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

Min-Hsuan Yeh, Max Kamachee, Seongheon Park, Yixuan Li

Researchers introduce HalluEntity, the first entity-level hallucination detection dataset derived from real LLM-generated text, to benchmark and understand detection methods. The study finds that context-aware uncertainty scores like Focus achieve better performance with an AUROC of 0.78, but all methods face challenges with sparse hallucinations, indicating the detection method is more critical than the LLM's scale.

Key insights

  • Basic uncertainty scores, such as Likelihood and Entropy, tend to over-predict hallucinations due to their lack of contextual awareness, resulting in low precision but high recall.
  • Context-aware methods like SAR and Focus demonstrate improved overall performance by integrating more contextual information, with Focus showing a better balance between precision and recall.
  • The choice of uncertainty scoring method is more influential on detection efficacy than the specific LLM family or its capacity, as increasing model size yielded only marginal performance improvements.

Summary by alphaXiv.

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