当纠正提示有害：在OWL~2~DL合规查询中LLM过度谨慎的修复提示设计
When Corrective Hints Hurt: Prompt Design in Reasoner-Guided Repair of LLM Overcaution on Entailed Negations under OWL~2~DL
AI总结：
研究发现提示设计对LLM在OWL~2~DL合规查询中的修复效果有显著影响，通过对比不同交互模式发现，基于推理器的修复方法在无提示时表现最佳，而带有提示的修复方法效果更差。
摘要：
我们报告了GPT-5.4在OWL~2~DL合规查询中的一种可重复错误模式：当推理器推导出的答案为'no'时，模型经常回答'unknown'。使用180个经过推理器审核的查询以及两个无关领域（保险和临床）的18个手工编制的测试查询，我们比较了四种交互模式：单次交互、三次通用'你错了'重试、三次基于推理器裁决的修复并带有开放世界假设（OWA）提示，以及无提示的修复。直接忠实度为43.9%（Wilson 95% CI [36.8,51.2]）；通用重试达到81.7%（[75.4,86.6]）；带有提示的裁决方法在67.2%（[60.1,73.7]）时表现更差；无提示的裁决方法达到97.8%（[94.4,99.1]）。所有配对比较在McNemar's精确检验和Bonferroni校正下仍显著（α=0.01；所有p<10^-5）。相同的指纹在测试查询中导致4/4个错误。我们的解释是受限的：提示框架可能比纠正内容更重要，推理器引导的包装器应被显式消解。
英文摘要：
We report a reproducible error pattern in GPT-5.4 on OWL~2~DL compliance queries: the model frequently answers ``unknown'' when the reasoner-entailed answer is ``no'' under \emph{FunctionalProperty} closure or class \emph{disjointness}. Using 180 reasoner-audited queries from a procedural expansion of the observed pattern plus 18 hand-authored held-out queries in two unrelated domains (insurance and clinical), we compare four interaction modes under matched query budget: single-shot, three rounds of generic ``you-are-wrong'' retry, three rounds of reasoner-verdict repair with an open-world-assumption (OWA) hint, and the same repair without the hint. Direct faithfulness is 43.9\,\% (Wilson 95\,\% CI $[36.8,51.2]$); generic retry reaches 81.7\,\% ($[75.4,86.6]$); the verdict-with-hint variant is \emph{worse} at 67.2\,\% ($[60.1,73.7]$); the verdict-only variant reaches 97.8\,\% ($[94.4,99.1]$). All pairwise comparisons remain significant under McNemar's exact test with Bonferroni correction ($α= 0.01$; all $p < 10^{-5}$). The same fingerprint accounts for 4/4 errors on the held-out queries. Our interpretation is bounded: prompt framing can matter more than corrective content, and reasoner-guided wrappers should be ablated explicitly.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
