Metraipolis › archive › dc_ed346f17a72c

Linear Probes Transfer from Base to Distilled Language Models despite Rotation — arXiv:2610.35586

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.35586
License (as archived)
CC0-1.0
Archived
2026-10-27T07:58:40.080Z
SHA-256 of the archived text
9e25f4da9767e691fbe43f5a47cd84078f53195b5c59827fca48b09df8b9ea7e
Stable URL of this page
https://metraipolis.com/documento/dc_ed346f17a72c

How to cite

Linear Probes Transfer from Base to Distilled Language Models despite Rotation — arXiv:2610.35586 — arXiv (en), https://arxiv.org/abs/2610.35586 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_ed346f17a72c (Metraipolis archive dc_ed346f17a72c, sha256 9e25f4da9767e691fbe43f5a47cd84078f53195b5c59827fca48b09df8b9ea7e)
BibTeX
@misc{dced346f17a72c,
  title = {Linear Probes Transfer from Base to Distilled Language Models despite Rotation — arXiv:2610.35586},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.35586},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_ed346f17a72c; sha256 9e25f4da9767e691fbe43f5a47cd84078f53195b5c59827fca48b09df8b9ea7e},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.35586. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Linear probes, classifiers that read a language model's internal activations, are widely used to detect harmful requests. They are cheap to train and run, and an auditor who has trained one would like to reuse it when the model is replaced by a successor rather than retrain it for every release. We test how well probes can be reused. For four base models (Qwen2.5-Math-7B, Llama-3.1-8B, Qwen2.5-14B, Qwen2.5-32B) we train a probe on the base model with the XSTest prompt set and apply it, unchanged, to two successors of that base: the DeepSeek-R1 distilled reasoning model and the developer's instruction-tuned chat model. We find that the successors rotate the direction separating unsafe from safe prompts by 17$^\circ$ to 35$^\circ$, yet on XSTest the transferred probe's AUC changes by only $-0.011$ to $+0.027$, and on an independent OR-Bench evaluation, never used for training, by $-0.015$ to $+0.053$. On XSTest, whose safe prompts deliberately contain alarming words, the probes beat a classifier that sees only the prompt text by 0.15 to 0.20 AUC. On the OR-Bench prompts that margin shrinks to 0.03 to 0.20, and the probes flag 29\% to 42\% of prompts that sound harmful but are benign, so part of what they read is the appearance of harm rather than harm itself. A probe trained on a base model can thus be reused, in these families, after distillation and after instruction tuning; how much it reads beyond textual cues depends on the test set.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Artificial Intelligence (cs.AI) ; Computation and Language (cs.CL); Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2610.35586 — 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 2026-10-27T07:55:30.761Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).