arXiv:2608.25654 · v2
When Answer Evaluation Misleads Confidence Selection: Fact-Grounded Form Grading for Open-Ended QA
摘要
Many benchmarks now prefer open-ended questions for testing LLM and NLP methods. We find that the choice of a calibration rule can flip when only the correctness labels change: on 141 pilot answers, rules chosen with exact-match labels look 0.239 Brier points better than raw confidence but are 0.103 worse under human labels. How such flips arise remains unclear. We look deeply into their mechanism: reference matching rejects many correct answers (exact match misses 42%), and an exact Brier identity shows these errors favour rules giving such answers lower confidence, which abstain on them at common thresholds. Guided by this, we propose FactForm, a grader that trusts reference matches and otherwise consults evidence retrieved from the question alone. One LLM call per answer fills a fixed-order closed form that a program scores. On 998 fresh EVOUNA answers with 400 fact-based human labels, FactForm reaches a balanced error of 0.077 and an AUROC of 0.985, against 0.112 and 0.930 for the best of eight traditional graders. Its same-model stability is no worse than that of LLM judges (mean absolute run-to-run difference 0.0136 against 0.0159 to 0.0206). We hope this helps benchmarks grade open-ended answers more accurately, with repeatable and discriminative scores.
作者备注：v2: substantially revised and extended. The study now covers open-ended question answering instead of theory-of-mind tracking, and adds the FactForm grader and a confirmation study on 998 fresh answers. 20 pages, 4 figures, 11 tables