Metraipolis › archive › dc_41ad7315bb61

Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models — arXiv:2609.12586

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.12586
License (as archived)
CC0-1.0
Archived
2026-10-16T03:02:49.587Z
SHA-256 of the archived text
95c1b7be0050a7a6335a20a63b102d21c2a2060abb26d41404723c75a2f97a04
Stable URL of this page
https://metraipolis.com/documento/dc_41ad7315bb61

How to cite

Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models — arXiv:2609.12586 — arXiv (en), https://arxiv.org/abs/2609.12586 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_41ad7315bb61 (Metraipolis archive dc_41ad7315bb61, sha256 95c1b7be0050a7a6335a20a63b102d21c2a2060abb26d41404723c75a2f97a04)
BibTeX
@misc{dc41ad7315bb61,
  title = {Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models — arXiv:2609.12586},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.12586},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_41ad7315bb61; sha256 95c1b7be0050a7a6335a20a63b102d21c2a2060abb26d41404723c75a2f97a04},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.12586. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

In this reproduction paper we investigate subliminal learning, a consequence of distillation where teacher models transmit behavioral preference traits through semantically unrelated data. The original paper explores two types of traits (animal preferences and misalignment), three data modalities (number sequences, code, and chain of thought), and several model families. We reproduce their experiments and extend the setup along three axes: new preference categories (actors and politicians), a new task (chess move generation), and an additional open-weight model (Ministral8B). We also run a controlled ablation on the numbers task's answer-space size (1-, 2-, and 3-digit sequences). We focus on open-weight models with accessible checkpoints on HuggingFace, since the original paper's GPT-4.x fine-tuning is no longer available. Our reproduction supports the original paper's claims, but our extensions show they are not universal as transmission strength varies across traits and tasks, and one model shows almost no effect at all.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-14 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2609.12586 — 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 2026-10-16T02:51:27.805Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-14?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).