Elamur Research⌕
☰
home / papers / 2026 / Can Activation Steering Generalize Across Languages? A Study on Syllogistic Reasoning in Language Models

Can Activation Steering Generalize Across Languages? A Study on Syllogistic Reasoning in Language Models

Paper · 2026 · Gabriele Maraia, Leonardo Ranaldi, Marco Valentino et al.

Summary

Large language models often judge a syllogism by whether its content sounds believable rather than whether the conclusion actually follows. This bias is known as the content effect. Earlier work showed that steering internal activations can reduce it, but only English had been tested. This paper asks whether steering assets built entirely from English data still help when the same arguments appear in other languages.

The authors translated an English syllogism corpus into other high-resource and low-resource languages and checked quality with round-trip back-translation. A logistic-regression probe trained on English activations predicts validity and selects a class-specific intervention. The intervention is either an additive shift toward a class centroid or a blend toward a content-free activation. That activation is predicted by networks that map concrete syllogisms onto placeholder forms such as "all X are Y".

Across six open models, English-derived steering raised formal reasoning accuracy in other languages by up to 36%. It barely affected perplexity and partly transferred to out-of-distribution logic tasks. No retraining is required, which makes this a cheap cross-lingual control mechanism.

Key points

  • Languages: The high-resource set is en, de, es, fr, it, ru and zh; the low-resource set is sw, bn and te. Translations passed a dual-pass integrity check based on back-translation.
  • Models: The study tested Qwen-2.5-7B, Gemma-2-9B, Mistral-7B, Llama-3.1-8B, EuroLLM-9B and Velvet-2B. Steering targets mean-pooled residual-stream activations from layers in the final third of each model.
  • Steering methods:
    • CEN-AST is additive and shifts activations toward a class centroid.
    • FFN-AST is a two-headed network that predicts the direction and magnitude of the target activation.
    • RRN-AST applies a rotation plus a residual translation in a PCA latent space.
    • All three are gated by an English-trained probe.
  • Main findings: CEN-AST was the most robust method. The blending methods worked, but their success depended on the internal geometry of each model. The probe's imperfect cross-lingual accuracy limits overall gains, so the authors introduced a Steering Efficiency metric.
  • Fluency check: Perplexity was measured on held-out Wikipedia text with steering active toward both the valid and the invalid class. Language modelling quality was only minimally affected. Code and datasets are released.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of EdinburghEdinburgh · United Kingdom359 works · 77,352 citations🎓University of SheffieldSheffield · United Kingdom113 works · 21,542 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 works

Cite

Cite as Gabriele Maraia, Leonardo Ranaldi, Marco Valentino and Fabio Massimo Zanzotto. 2026. Can Activation Steering Generalize Across Languages? A Study on Syllogistic Reasoning in Language Models. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-long.125

@inproceedings{maraia2026can,
  title = {{Can Activation Steering Generalize Across Languages? A Study on Syllogistic Reasoning in Language Models}},
  author = {Gabriele Maraia, Leonardo Ranaldi, Marco Valentino and Fabio Massimo Zanzotto},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.125},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.125}
}

References

28 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pombal, Nicolas Boizard, Manuel Faysse, Pierre Colombo, François Yvon, Barry Haddow, José G. C. de Souza, Alexandra Birch, and André F. T. Martins. 2025. Eurollm-9b: Technical report. Preprint, arXiv:2506.04079.
  2. AI@Meta. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  3. Almawave. 2024. Velvet-2b: A 2b-parameter open large language model. https://huggingface.co/ Almawave/Velvet-2B. Accessed: 2025-09-06. ↗
  4. Leonardo Bertolazzi, Albert Gatt, and Raffaella Bernardi. 2024. A systematic analysis of large language models as soft reasoners: The case of syllogistic inferences. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 13882–13905, Miami, Florida, USA. Ass… find
  5. Peter Clark, Oyvind Tafjord, and Kyle Richardson. 2020. Transformers as soft reasoners over language. Preprint, arXiv:2002.05867.
  6. Boyi Deng, Yu Wan, Baosong Yang, Yidan Zhang, and Fuli Feng. 2025. Unveiling language-specific features in large language models via sparse autoencoders. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4563–4608, Vienna, Austr…
  7. Tiwalayo Eisape, Michael Tessler, Ishita Dasgupta, Fei Sha, Sjoerd Steenkiste, and Tal Linzen. 2024. A systematic comparison of syllogistic reasoning in humans and language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: H… find
  8. J. St. B. T. Evans, Julie L. Barston, and Paul Pollard. 1983. On the conflict between logic and belief in syllogistic reasoning. Memory amp; Cognition, 11(3):295–306. find
  9. Gemma Team and 1 others. 2024. Gemma 2: Improving open language models at a practical size. Preprint, arXiv:2408.00118. find
  10. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix,… find
  11. Kentaro Ozeki, Risako Ando, Takanobu Morishita, Hi-rohiko Abe, Koji Mineshima, and Mitsuhiro Okada. 2024. Exploring reasoning biases in large language models through syllogism: Insights from the NeuBAROCO dataset. In Findings of the Association for Computational Linguistics: ACL 2024, pages 16063–16… find
  12. Giulia Pucci and Leonardo Ranaldi. 2025. Advancing oversight reasoning across languages for audit sycophantic behaviour via X-agent. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 12949–12965, Suzhou, China. Association for Computational Linguistics. find
  13. Libo Qin, Qiguang Chen, Fuxuan Wei, Shijue Huang, and Wanxiang Che. 2023. Cross-lingual prompting: Improving zero-shot chain-of-thought reasoning across languages. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 2695–2709, Singapore. Association for C…
  14. Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, and 25 others. 2025. Qwen2.5 technical report. Preprint, arXiv:2412.15115. find
  15. Leonardo Ranaldi and Andre Freitas. 2024. Aligning large and small language models via chain-of-thought reasoning. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1812–1827, St. Julian’s, Malta. Association…
  16. Leonardo Ranaldi and Giulia Pucci. 2025. Multilingual reasoning via self-training. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 11566–11582, Albuquerque, New M… find
  17. Geonhee Kim, Marco Valentino, and Andre Freitas. 2025. Reasoning circuits in language models: A mechanistic interpretation of syllogistic inference. In Findings of the Association for Computational Linguistics: ACL 2025, pages 10074–10095, Vienna, Austria. Association for Computational Linguistics. find
  18. Leonardo Ranaldi, Giulia Pucci, Federico Ranaldi, Elena Sofia Ruzzetti, and Fabio Massimo Zanzotto. 2024. A tree-of-thoughts to broaden multi-step reasoning across languages. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 1229–1241, Mexico City, Mexico. Association f… find
  19. Pedro Henrique Martins, João Alves, Patrick Fernandes, Nuno M. Guerreiro, Ricardo Rei, Amin Farajian, Mateusz Klimaszewski, Duarte M. Alves, José Leonardo Ranaldi, Federico Ranaldi, and Giulia Pucci. 2025a. R2-MultiOmnia: Leading multilingual multimodal reasoning via self-training. In Proceedings of…
  20. Leonardo Ranaldi, Federico Ranaldi, Fabio Massimo Zanzotto, Barry Haddow, and Alexandra Birch. 2025b. Improving multilingual retrieval-augmented language models through dialectic reasoning argumentations. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, page…
  21. Leonardo Ranaldi, Marco Valentino, and Andre Freitas. 2025c. Improving chain-of-thought reasoning via quasi-symbolic abstractions. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 17222–17240, Vienna, Austria. Association for C…
  22. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steering llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15504–15522, Bangkok, Thaila… find
  23. Samuel Soo, Chen Guang, Wesley Teng, Chandrasekaran Balaganesh, Tan Guoxian, and Yan Ming. 2025. Interpretable steering of large language models with feature guided activation additions. Preprint, arXiv:2501.09929. — [arXiv:2501.09929](https://arxiv.org/abs/2501.09929) find
  24. Niklas Stoehr, Kevin Du, Vésteinn Snæbjarnarson, Robert West, Ryan Cotterell, and Aaron Schein. 2024. Activation scaling for steering and interpreting language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 8189–8200, Miami, Florida, USA. Association for Comp… find
  25. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2024. Steering language models with activation engineering. Preprint, arXiv:2308.10248.
  26. Marco Valentino, Geonhee Kim, Dhairya Dalal, Zhixue Zhao, and André Freitas. 2025. Mitigating content effects on reasoning in language models through fine-grained activation steering. ArXiv, abs/2505.12189. — [arXiv:2505.12189](https://arxiv.org/abs/2505.12189) find
  27. Magdalena Wysocka, Danilo Carvalho, Oskar Wysocki, Marco Valentino, and Andre Freitas. 2025. SylloBio-NLI: Evaluating large language models on biomedical syllogistic reasoning. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguisti… find
  28. Fabio Massimo Zanzotto, Elena Sofia Ruzzetti, Giancarlo A. Xompero, Leonardo Ranaldi, Davide Venditti, Federico Ranaldi, Cristina Giannone, Andrea Favalli, and Raniero Romagnoli. 2025. Position paper: MeMo: Towards language models with associative memory mechanisms. In Findings of the Association fo… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/maraia-2026-can-activation-steering-generalize-across/