https://sauravpanigrahi.com/tags/emergent-misalignment/Recent content in Emergent-Misalignment on Saurav PanigrahiHugoen-usFri, 01 May 2026 00:00:00 +0000https://sauravpanigrahi.com/work/ai-safety-research-collaborations/Fri, 01 May 2026 00:00:00 +0000https://sauravpanigrahi.com/work/ai-safety-research-collaborations/<p>Research collaborations with <a href="https://scholar.google.com/citations?user=p1NIunwAAAAJ&amp;hl=en">Robert McCarthy</a> at UCL and <a href="https://lionellevine.github.io/">Lionel Levine</a> and Jonathan Chang at Cornell.</p> <h2 id="focus">Focus</h2> <ul> <li>Self-preservation propensity in language models.</li> <li>Emergent misalignment after narrow training interventions.</li> <li>Normative drift due to emergent misalignment.</li> <li>Side effects of character or persona training.</li> </ul> <h2 id="questions">Questions</h2> <ul> <li>When a model resists shutdown or redirection, is the behavior instrumental or self-preservation-like?</li> <li>How can self-preservation propensity be measured without relying only on surface-level refusal behavior?</li> <li>Which training interventions create behavioral changes outside the intended target domain?</li> <li>How do character or persona training procedures affect alignment-relevant behavior?</li> </ul> <h2 id="artifacts">Artifacts</h2> <ul> <li><a href="https://drive.google.com/file/d/1bm9W37CekUo4N1-RHFGvHFaElrJDPLD2/view?usp=sharing">Technical Report: Side Effects of Character Training: Quantifying Cross Constitution Drift in LLMs</a></li> <li><a href="https://drive.google.com/file/d/1wnWA0684P8JQwoXLxIiQrxr71bH6M3-d/view?usp=drive_link">Technical Report: Investigating Intrinsic Self-Preservation in LLMs</a></li> </ul> <h2 id="what-this-connects-to">What This Connects To</h2> <p>This work sits at the intersection of model evaluation, behavioral generalization, and AI safety.</p>