Metraipolis › archive › dc_bacab8fccf74

"Not" is Not Lost in Translation: A Causal Analysis of Negation Representations in Multilingual LLMs — arXiv:2610.32281

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.32281
License (as archived)
CC0-1.0
Archived
2026-10-26T19:51:51.757Z
SHA-256 of the archived text
ab10e2d06ec8b75f59e6a9a8579b1403fb75e11de1979710ae9797b03f83b357
Stable URL of this page
https://metraipolis.com/documento/dc_bacab8fccf74

How to cite

"Not" is Not Lost in Translation: A Causal Analysis of Negation Representations in Multilingual LLMs — arXiv:2610.32281 — arXiv (en), https://arxiv.org/abs/2610.32281 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_bacab8fccf74 (Metraipolis archive dc_bacab8fccf74, sha256 ab10e2d06ec8b75f59e6a9a8579b1403fb75e11de1979710ae9797b03f83b357)
BibTeX
@misc{dcbacab8fccf74,
  title = {"Not" is Not Lost in Translation: A Causal Analysis of Negation Representations in Multilingual LLMs — arXiv:2610.32281},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.32281},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_bacab8fccf74; sha256 ab10e2d06ec8b75f59e6a9a8579b1403fb75e11de1979710ae9797b03f83b357},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.32281. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Negation occurs in every language, yet its surface realization differs drastically. Despite this, multilingual large language models (LLMs) translate negation robustly, raising the question of whether they share an underlying representation of negation across languages. We study this question in six open models across English, French, Vietnamese, and Chinese, using causal interventions on a synthetic contrastive dataset and a negation translation dataset derived from FLORES-200. Across models, we find strong evidence that negation representations are causally controllable and largely shared across languages. In particular, we show that a contrastive activation direction extracted from one language can induce or remove negation in the others, and directions extracted from different languages are substantially aligned. We further find that negation is represented in sparse features: a single sparse autoencoder (SAE) feature per model suffices to control polarity across languages, while responding selectively to negated inputs. Finally, both dense and sparse interventions generalize to four additional languages and FLORES-derived sentences. We release data and code to support future work: this https URL .

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2610.32281 — 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 2026-10-26T19:45:33.405Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).