| title | See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs |
|---|---|
| authors | Weiqiao Que, Ruizhe Li, Chengyu Wang, Dakan Wang, Emine Yilmaz, Xiaofeng He |
| date | 2026-09-28 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2609.34970 |
| affiliation | University of Birmingham |
| upvotes | 4 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | Safety-aligned LLMs can exhibit emergent misalignment (EM): narrow domain adaptation unexpectedly triggers catastrophic safety failures across unrelated domains. Prior static analyses leave training dynamics unmapped, while existing defenses rely on heuristics that degrade utility. We present a dynamic, second-order geometric study of EM. Tracking training trajectories reveals that directional Hessian curvature concentrates sharply on semantic pivot tokens. Grassmannian projections show that, in most settings, harmful-safe gap widens mainly because safe-gradient overlap declines. Leveraging these insights, we introduce a parameter-level Geometric Mitigation Framework that orthogonally projects empirical harmful gradient subspace out of parameter updates. On Qwen2.5-14B-IT, our defense suppresses free-generation EM by up to 80.0%; across the other three of four open-weight instruction-based model families (3B--20B), where single-layer behavioral EM is already near zero, teacher-forced evaluation shows same harmful subspace controls the conditional support of frozen EM responses. Crucially, these diagnostics unmask the illusion of behavioral safety: the same subspace remains measurable and steerable in models where behavioral EM is near zero. Code: https://github.com/WeiqiaoQUE/mechanistic-emergent-misalignment. |
cite
@misc{que2026see,
title={See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs},
author={Weiqiao Que and Ruizhe Li and Chengyu Wang and Dakan Wang and Emine Yilmaz and Xiaofeng He},
year={2026},
eprint={2609.34970},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2609.34970}
}