Metraipolis › archive › dc_14e488a2be21

Suppressing Misaligned Behaviors in Language Models, Durably — arXiv:2611.05358

Source: arXiv (en)
Original: https://arxiv.org/abs/2611.05358
License (as archived): CC0-1.0
Archived: 2026-11-04T18:37:39.892Z
SHA-256 of the archived text: c44151e98816005c6ef144dc138a6481460af11f7d725079940dc118ad201870
Stable URL of this page: https://metraipolis.com/documento/dc_14e488a2be21

How to cite

Suppressing Misaligned Behaviors in Language Models, Durably — arXiv:2611.05358 — arXiv (en), https://arxiv.org/abs/2611.05358 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_14e488a2be21 (Metraipolis archive dc_14e488a2be21, sha256 c44151e98816005c6ef144dc138a6481460af11f7d725079940dc118ad201870)

BibTeX

@misc{dc14e488a2be21,
  title = {Suppressing Misaligned Behaviors in Language Models, Durably — arXiv:2611.05358},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2611.05358},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_14e488a2be21; sha256 c44151e98816005c6ef144dc138a6481460af11f7d725079940dc118ad201870},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.05358. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Undesired model behaviors can emerge or re-emerge at any stage of training. Once observed, how effectively can we suppress them? To study this question, we first construct two model organisms--language models deliberately trained to exhibit a specific misalignment--that differ in what triggers the behavior. The narrow-trigger organism misbehaves only when the prompt contains a specific cue: it cheats on multiple-choice questions by following a planted hint that points to a wrong answer, and behaves normally otherwise. The broad-trigger organism has no such cue: after fine-tuning on insecure code, it gives harmful responses across a wide range of open-ended prompts. We then formalize post-hoc behavioral suppression and evaluate methods along three criteria: immediate suppression of the target behavior, durability under further training, and preservation of general capability. We compare standard methods--SFT, GRPO, and gradient-ascent unlearning--along these dimensions. Building on prior work on alignment shallowness, we hypothesize that more durable suppression requires the ability to recover from a wider range of misalignment states. We introduce Prefix-Expanded Adversarial Reinforcement Learning (PEARL), a GRPO variant that augments rollouts with continuations from intermediate states sampled along cached organism trajectories. Across both organism settings and two model families (Qwen3-4B, GPT-OSS-20B), PEARL achieves strong immediate suppression while preserving task accuracy, and remains durable under further fine-tuning on either the organism's data or unrelated data. It obtains the greatest benefit when the undesirable behavior is broad and cannot be addressed through targeted SFT. Together, our framework and method offer a step toward more durable post-hoc removal of undesired behaviors.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI); Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2611.05358 — 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 2026-11-04T18:32:37.498Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).