Metraipolis › archive › dc_fd3b556ce425

Strengthening Target-Language Features: SAE-Based Steering for Multilingual Inference — arXiv:2608.04904

Source
arXiv (en)
Original
https://arxiv.org/abs/2608.04904
License (as archived)
CC0-1.0
Archived
2026-10-31T11:25:19.048Z
SHA-256 of the archived text
85aa5be1ee7c1ebcfd74b8f40a5660f267be6d171515c349853b4ba5a6cb4595
Stable URL of this page
https://metraipolis.com/documento/dc_fd3b556ce425

How to cite

Strengthening Target-Language Features: SAE-Based Steering for Multilingual Inference — arXiv:2608.04904 — arXiv (en), https://arxiv.org/abs/2608.04904 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_fd3b556ce425 (Metraipolis archive dc_fd3b556ce425, sha256 85aa5be1ee7c1ebcfd74b8f40a5660f267be6d171515c349853b4ba5a6cb4595)
BibTeX
@misc{dcfd3b556ce425,
  title = {Strengthening Target-Language Features: SAE-Based Steering for Multilingual Inference — arXiv:2608.04904},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.04904},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_fd3b556ce425; sha256 85aa5be1ee7c1ebcfd74b8f40a5660f267be6d171515c349853b4ba5a6cb4595},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.04904. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Multilingual large language models exhibit substantial performance differences across languages, while existing adaptation methods often require parameter updates and considerable multilingual training data. We propose an inference-time multilingual steering method that uses pretrained sparse autoencoders to identify and strengthen target-language-related features. Using multilingual parallel sentences, we compare SAE activations across languages and select a small number of layer-specific features associated with each target language. These features are decoded into steering signals and injected into the model's hidden states without additional training. Experiments with Gemma-3-12B-it show macro-average accuracy improvements of 11.28 percentage points on XCOPA, 4.11 points on XNLI, and 7.84 points on PAWS-X. Additional evaluation with Qwen2.5-7B-Instruct yields a macro-average improvement of 3.33 percentage points on XCOPA.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-06 in cs.CL).
Categorie arXiv: Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2608.04904 — 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 2026-10-31T11:13:23.191Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-08-06?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).