Metraipolis › archive › dc_207e4334e23a

Selecting The Most Informative Tokens in Natural Language Autoencoders — arXiv:2609.37040

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.37040
License (as archived)
CC0-1.0
Archived
2026-10-07T16:26:45.569Z
SHA-256 of the archived text
b2a2eb423c7054510198ed87aae480e51dbde552c7f2495ed7709f41df54659a
Stable URL of this page
https://metraipolis.com/documento/dc_207e4334e23a

How to cite

Selecting The Most Informative Tokens in Natural Language Autoencoders — arXiv:2609.37040 — arXiv (en), https://arxiv.org/abs/2609.37040 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_207e4334e23a (Metraipolis archive dc_207e4334e23a, sha256 b2a2eb423c7054510198ed87aae480e51dbde552c7f2495ed7709f41df54659a)
BibTeX
@misc{dc207e4334e23a,
  title = {Selecting The Most Informative Tokens in Natural Language Autoencoders — arXiv:2609.37040},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.37040},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_207e4334e23a; sha256 b2a2eb423c7054510198ed87aae480e51dbde552c7f2495ed7709f41df54659a},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.37040. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Natural language autoencoders translate a language model's internal activations into readable explanations. Explaining every token position is costly. Which positions should an auditor inspect to understand a potential threat? We study this question across $4.7$ million explanations on prompt injection and concealment. We compare signals from model computation with a ranker trained only on chat structure. Chat structure usually selects more relevant explanations than the computational signals, without requiring a model forward pass for position selection. On three of four datasets, explaining just $5\%$ of positions retains nearly all of the success rate from explaining every position, where success means obtaining an explanation about the threat. The benefit varies with the audit task. We also show that pretrained verbalizers recover words that models have learned to conceal through fine-tuning, without additional verbalizer training. These results identify where auditors can concentrate explanation generation and show that useful explanations can extend beyond the model a verbalizer was trained to describe.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-30 in cs.AI).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2609.37040 — 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 2026-10-07T15:54:53.932Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-30?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).