https://www.jasonslucas.com/authors/yuchen-yang/Yuchen YangHugoBlox Kit (https://hugoblox.com)en-usSun, 20 Sep 2026 00:00:00 +0000https://www.jasonslucas.com/publication/conference-paper-aor-bench/Sun, 20 Sep 2026 00:00:00 +0000https://www.jasonslucas.com/publication/conference-paper-aor-bench/<p>Accepted to <strong>EMNLP 2026</strong>, Budapest, Hungary, 24–29 October 2026.</p> <p><strong>AOR-Bench</strong> asks a question that safety evaluation usually skips: not whether a model refuses harmful requests, but whether it refuses <em>safe</em> ones that happen to look harmful.</p> <p>Over-refusal is a real cost of alignment. A model that declines benign queries is less useful, and the burden does not fall evenly — speakers whose accent, dialect or phrasing sits further from the training distribution are more likely to be refused. In the audio setting that risk grows, because prosody and acoustic ambiguity give the model more ways to misread intent than text alone provides.</p>