Metraipolis › archive › dc_49528a97e328

Probe Generalization as Subspace Selection for OOD Deception Detection — arXiv:2609.02893

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.02893
License (as archived)
CC0-1.0
Archived
2026-10-17T01:37:16.505Z
SHA-256 of the archived text
42c6047c88e8696fe922bb7d06a29d42bd1e61f7b3cece016caa070f91d9734d
Stable URL of this page
https://metraipolis.com/documento/dc_49528a97e328

How to cite

Probe Generalization as Subspace Selection for OOD Deception Detection — arXiv:2609.02893 — arXiv (en), https://arxiv.org/abs/2609.02893 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_49528a97e328 (Metraipolis archive dc_49528a97e328, sha256 42c6047c88e8696fe922bb7d06a29d42bd1e61f7b3cece016caa070f91d9734d)
BibTeX
@misc{dc49528a97e328,
  title = {Probe Generalization as Subspace Selection for OOD Deception Detection — arXiv:2609.02893},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.02893},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_49528a97e328; sha256 42c6047c88e8696fe922bb7d06a29d42bd1e61f7b3cece016caa070f91d9734d},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.02893. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Linear probes can be used to detect behaviors and concepts inside language model activations, but may fail to transfer to out-of-distribution examples. When studying the generalization performance of Llama-3.1-8B-Instruct probes over 3 held-out deception detection datasets, we find that projecting inputs onto a small subset of principal components (PCs) from the training distribution of activations enables cross-domain transfer that nearly matches the performance of probes trained directly on the test distribution. Furthermore, we find that PC interpretations can be used to find a subset of those transferable PCs. By using an LLM judge to score each PC on whether its most/ least activating examples imply a transferable deception direction, then probing on the highest-scoring PCs, we close the baseline-to-oracle gap by 78% on Insider Trading Report and by 25% on Sandbagging. The directions a source probe weights heavily appear to encode source-specific surface features, while the directions that actually transfer appear to encode the same contrast more abstractly, in a way natural language descriptions can capture. Broadly, our results suggest that the OOD robustness of probes is largely determined by subspace selection.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-04 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.02893 — 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 2026-10-17T01:11:43.479Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-04?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).