Metraipolis › archive › dc_56da978da793

Inherited or Learned? Backdoor Selectivity in Distilled Language Models — arXiv:2611.00748

Source: arXiv (en)
Original: https://arxiv.org/abs/2611.00748
License (as archived): CC0-1.0
Archived: 2026-11-03T11:01:27.589Z
SHA-256 of the archived text: 841a873007bf0050cbc50fe6e4f94be4a631ad4cceed650cea82b1a385d514e9
Stable URL of this page: https://metraipolis.com/documento/dc_56da978da793

How to cite

Inherited or Learned? Backdoor Selectivity in Distilled Language Models — arXiv:2611.00748 — arXiv (en), https://arxiv.org/abs/2611.00748 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_56da978da793 (Metraipolis archive dc_56da978da793, sha256 841a873007bf0050cbc50fe6e4f94be4a631ad4cceed650cea82b1a385d514e9)

BibTeX

@misc{dc56da978da793,
  title = {Inherited or Learned? Backdoor Selectivity in Distilled Language Models — arXiv:2611.00748},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2611.00748},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_56da978da793; sha256 841a873007bf0050cbc50fe6e4f94be4a631ad4cceed650cea82b1a385d514e9},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.00748. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Distillation is a routine step in building small language models, and a backdoored teacher can pass its backdoor to the student. We study what a student inherits when the backdoor is a rare string that triggers a fixed refusal. Keeping the trigger fixed, we vary only where it appears, at a random word boundary or appended to the instruction, and compare teachers and students on the exact trigger, six control strings and clean inputs. Across Qwen2.5, Llama 3.2 and Gemma 3, with 5% of transfer prompts triggered, randomly inserted triggers become less selective after distillation in all nine teacher-student pipelines, while appended triggers gain selectivity in Qwen and Llama and change little in Gemma. Students trained on the same prompts without a teacher show that much of this is not inherited. Llama and Gemma students learn an appended backdoor from the 150 poisoned examples alone, and only Qwen's 0.5B student needs the teacher. The data also sets where the trigger works. Gemma teachers respond to it anywhere, but their students only when it is appended. Audits should compare teacher and student on the same inputs, probe several trigger positions and include a student trained without the teacher.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2611.00748 — 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 2026-11-03T10:59:45.737Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).