Reproduce “The Labeling Problem in Hallucination Detection Benchmarks: An Empirical Evaluation” 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
Jorma Valjakka, Juhani Kivimäki, Juha Mylläri, Jukka K. Nurminen
The authors examine how automated labelers assess hallucinations in open-domain question answering when references are short. They compare reference faithfulness with factual correctness and test judges against human annotations, including a prompt ablation. For Mistral-7B answers, GPT-5-mini’s Cohen’s κ rises from 0.194 to 0.626; this is one comparison on a limited English QA audit set.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.