Metraipolis › archive › dc_2b11bbe28d3f

Probing and Steering for Safety-Relevant Behaviours in Large Language Models — arXiv:2610.30363

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.30363
License (as archived)
CC0-1.0
Archived
2026-10-25T08:33:51.579Z
SHA-256 of the archived text
04ec43d450c4ca91f9ea01bc1d2fdd03dc3f26f694e022d1f5d2dfe22b0419cd
Stable URL of this page
https://metraipolis.com/documento/dc_2b11bbe28d3f

How to cite

Probing and Steering for Safety-Relevant Behaviours in Large Language Models — arXiv:2610.30363 — arXiv (en), https://arxiv.org/abs/2610.30363 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_2b11bbe28d3f (Metraipolis archive dc_2b11bbe28d3f, sha256 04ec43d450c4ca91f9ea01bc1d2fdd03dc3f26f694e022d1f5d2dfe22b0419cd)
BibTeX
@misc{dc2b11bbe28d3f,
  title = {Probing and Steering for Safety-Relevant Behaviours in Large Language Models — arXiv:2610.30363},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.30363},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_2b11bbe28d3f; sha256 04ec43d450c4ca91f9ea01bc1d2fdd03dc3f26f694e022d1f5d2dfe22b0419cd},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.30363. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Activation probes are computed directions in a language model's residual stream, used both to detect safety-relevant behaviours and, by adding their direction back into the stream, to steer them. We explore how the main design choices behind such probes interact, within one pipeline, on one model (Llama-3.1-8B-Instruct), and across three behaviours: deception, high-stakes awareness and sycophancy. The pipeline crosses two estimators (difference-of-means and logistic regression) with three token-pooling schemes (last, mean and learned attention), selects appropriate layers to detect or steer a behaviour, and evaluates every direction twice: as a detector of the behaviour, and as a steering vector to push outputs towards the behaviour. Prior work read together suggests a read/write dissociation, in which logistic regression reads (detects) best but difference-of-means directions write (steer) best. We find that it holds for deception and high-stakes but not for sycophancy, where the pattern reverses. The reversal appears to follow partially from the data rather than from the estimators. The best pooling technique and the layer choice depend on the behaviour and data, and the layer that steers best can sit well before the layer that detects best. The right steering metric to use between changing a model's perception and eliciting behaviour depends on the task - and the two can disagree outright. The exported probes and directions power an interactive interface through which a user can score and steer text directly.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Machine Learning (cs.LG) ; Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY).
Pagina del lavoro: https://arxiv.org/abs/2610.30363 — 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 2026-10-25T08:31:13.143Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).