http://xzhou98.github.io/tags/neurips/NeurIPSHugoBlox Kit (https://hugoblox.com)en-usMon, 05 Oct 2026 00:00:00 +0000http://xzhou98.github.io/media/icon_hu_da05098ef60dc2e7.pnghttp://xzhou98.github.io/tags/neurips/http://xzhou98.github.io/blog/neurips-26/Mon, 05 Oct 2026 00:00:00 +0000http://xzhou98.github.io/blog/neurips-26/<p>Our paper <strong>&ldquo;Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models&rdquo;</strong> has been accepted to <strong>NeurIPS 2026</strong>. This is joint work with Saleh Zare Zade, Rafi Ibn Sultan, Alexander Kotov, and my advisor Dongxiao Zhu.</p> <p>Large reasoning models are typically trained with rewards on their final answers, leaving the reasoning trace largely unsupervised. We show this produces <em>deceptive safety alignment</em>: a safe-looking final answer that sits on top of unsafe reasoning. We introduce <strong>DSAR</strong>, a metric that quantifies this inconsistency, and <strong>SARA</strong>, an RL method that rewards safety-aware reasoning as well as safe answers — reducing deceptive alignment under both standard prompting and prefilling attacks while preserving helpfulness.</p> <p> · · </p>