Reproduce “Rethinking Verbalized Confidence for LLM-as-a-Judge: A Compatibility Shift on Post-2025 Proprietary 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
Yu-Chung Hsiao
The authors test a single-call LLM judge that returns a binary verdict and confidence score, adding an overconfidence advisory and a request to argue both sides before deciding. On SummEval, verbalized confidence becomes more robust to task subjectivity than logprob-based G-Eval in matched GPT-5 releases, but this comparison is limited to the GPT family and this task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.