Metraipolis › archive › dc_f2cc8714300b

DeShortcut-Align: Decoupling Spurious Shortcuts for Robust Safety Alignment in Large Reasoning Models — arXiv:2609.34896

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.34896
License (as archived)
CC0-1.0
Archived
2026-10-07T20:58:18.126Z
SHA-256 of the archived text
bb0c09014ec325c633823672772e6c7a8ca55593b405d8c41fb3a398811eec87
Stable URL of this page
https://metraipolis.com/documento/dc_f2cc8714300b

How to cite

DeShortcut-Align: Decoupling Spurious Shortcuts for Robust Safety Alignment in Large Reasoning Models — arXiv:2609.34896 — arXiv (en), https://arxiv.org/abs/2609.34896 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_f2cc8714300b (Metraipolis archive dc_f2cc8714300b, sha256 bb0c09014ec325c633823672772e6c7a8ca55593b405d8c41fb3a398811eec87)
BibTeX
@misc{dcf2cc8714300b,
  title = {DeShortcut-Align: Decoupling Spurious Shortcuts for Robust Safety Alignment in Large Reasoning Models — arXiv:2609.34896},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.34896},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_f2cc8714300b; sha256 bb0c09014ec325c633823672772e6c7a8ca55593b405d8c41fb3a398811eec87},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.34896. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Safety alignment of large reasoning models (LRMs) via supervised fine-tuning (SFT) and reinforcement learning (RL) often yields near-perfect safety scores, yet this apparent success comes at the cost of severe over-refusal and degraded general capabilities. Through systematic empirical analysis, we find that these failures are closely associated with the learning of spurious shortcuts rather than robust intent-sensitive safety evaluation. Specifically, we identify two dominant shortcuts: formatting shortcuts, where refusal behaviors are overly bound to structural prompt templates that frequently appear in safety alignment corpora; and lexical shortcuts, where sensitive keywords reflexively trigger refusals on benign queries. To mitigate reliance on these shortcuts, we propose DeShortcut-Align, a shortcut-decoupling alignment framework that reduces dependence on superficial cues. DeShortcut-Align operates across three coordinated stages: (1) Refusal Sensitivity Attribution, which masks input tokens to quantify their impact on the final refusal response distribution; (2) Attribution-Guided Contrastive Augmentation, which constructs benign contrastive samples using high-sensitivity tokens to mitigate lexical shortcuts; and (3) Counterfactual Consistency Regularization, which constructs template-ablated states via attention blinding to enforce decision consistency across SFT and RL, mitigating formatting shortcut dependence. Experiments on 7B and 14B models demonstrate that DeShortcut-Align significantly improves robustness against template-stripping bypass attacks (reducing performance drops by up to 72%), substantially reduces over-refusal by over 58%, and better preserves general-purpose reasoning capabilities, thereby mitigating the alignment tax commonly observed in safety training.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-29 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2609.34896 — 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 2026-10-07T20:29:01.708Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-29?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).