Metraipolis › archive › dc_a715ce4dbd2d

See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs — arXiv:2609.34970

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.34970
License (as archived)
CC0-1.0
Archived
2026-10-07T21:54:17.965Z
SHA-256 of the archived text
6ca9ee83d85634ad3e22ef2d442ced17bfd71865a46cb6b2dafdf68d1cb1ee07
Stable URL of this page
https://metraipolis.com/documento/dc_a715ce4dbd2d

How to cite

See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs — arXiv:2609.34970 — arXiv (en), https://arxiv.org/abs/2609.34970 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_a715ce4dbd2d (Metraipolis archive dc_a715ce4dbd2d, sha256 6ca9ee83d85634ad3e22ef2d442ced17bfd71865a46cb6b2dafdf68d1cb1ee07)
BibTeX
@misc{dca715ce4dbd2d,
  title = {See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs — arXiv:2609.34970},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.34970},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_a715ce4dbd2d; sha256 6ca9ee83d85634ad3e22ef2d442ced17bfd71865a46cb6b2dafdf68d1cb1ee07},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.34970. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

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: this https URL .

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-29 in cs.CL).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.34970 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-07T21:27:08.836Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-29?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).