Reproduce “DirEAG: Dirichlet Evidence Aggregation for Calibrating Verbalized Confidence in Mathematical Reasoning” 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
Haorui Xu, Yuzhou Zhu, Liyuan Gao
DirEAG learns to combine answer-and-confidence reports from five prompts by aligning prompt-specific confidence scales and treating them as evidence over candidate answers plus a null state for missing answers. A final calibrator estimates correctness probability. On Gemma-2-9B-it and SVAMP, it achieves ECE of 0.0418, versus 0.0954 for SteerConf, with accuracy 0.8230 versus 0.8070; this result is representative, not universal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.