---
title: "Can Dialects Be Steered Like Languages? — Elozeiri 2026"
description: "Arabic LLMs are trained mostly on Modern Standard Arabic (MSA), so they tend to answer in MSA even when a user writes in a dialect."
authors: ["Kareem Elozeiri", "Mervat Abassy", "Omar Kallas", "Fahim Dalvi", "Preslav Nakov", "Kentaro Inui", "Nadir Durrani"]
year: 2026
venue: "arXiv:2607.03936"
source: "https://arxiv.org/abs/2607.03936"
doi: "10.48550/arxiv.2607.03936"
arxiv: "2607.03936"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/elozeiri-2026-dialects-steered-like-languages-sparse/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/elozeiri-2026-dialects-steered-like-languages-sparse/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/2/) / Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMs

# Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMs

**Authors:** [Kareem Elozeiri](https://md.research.elamur.ai/free/en/authors/kareem-elozeiri/), [Mervat Abassy](https://md.research.elamur.ai/free/en/authors/mervat-abassy/), [Omar Kallas](https://md.research.elamur.ai/free/en/authors/omar-kallas/), [Fahim Dalvi](https://md.research.elamur.ai/free/en/authors/fahim-dalvi/), [Preslav Nakov](https://md.research.elamur.ai/free/en/authors/preslav-nakov/), [Kentaro Inui](https://md.research.elamur.ai/free/en/authors/kentaro-inui/), [Nadir Durrani](https://md.research.elamur.ai/free/en/authors/nadir-durrani/) · **Year:** 2026 · **Venue:** arXiv:2607.03936 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2607.03936) · [PDF](https://arxiv.org/pdf/2607.03936) · [DOI 10.48550/arxiv.2607.03936](https://doi.org/10.48550/arxiv.2607.03936) · [arXiv 2607.03936](https://arxiv.org/abs/2607.03936) · [OpenAlex](https://openalex.org/W7167586886)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14) · 1 in our library, of the 140884 papers whose bibliographies we read

## Summary
Arabic LLMs are trained mostly on Modern Standard Arabic (MSA), so they tend to answer in MSA even when a user writes in a dialect. The authors ask where dialect identity lives inside such models and whether it can be pushed at inference time without any fine-tuning. They compare two interventions on frozen weights. The first finds dialect-selective MLP neurons with a LAPE (entropy-based) criterion and rescales them: it boosts the target dialect and damps MSA and competitor dialects. The second builds a mean dialect-minus-MSA activation vector from parallel MADAR sentence pairs and adds it to the residual stream while decoding. The models are ALLaM and Fanar, evaluated on Egyptian, Moroccan, Levantine and Gulf varieties. The selected neurons do exist and cluster in late layers, but they explain only a small slice of the overall dialect direction. As a result, neuron steering can strengthen a dialect that the prompt already uses, but it cannot create one from an MSA prompt. Vector steering works in both cases. The conclusion is that dialect control in these models depends mainly on distributed representations rather than localized neurons.

## Key points
- Models: ALLaM-7B-Instruct-preview (32 layers) and Fanar-1-9B-Instruct (42 layers), both instruction-tuned for Arabic.
- Data: MADAR parallel corpus.
  - Cairo, Rabat, Beirut and Doha each supply 12,000 sentence pairs; Riyadh and Aleppo supply only 2,000.
  - Ablations use Egyptian and Moroccan.
  - Beirut and Aleppo are scored on the benchmark's Syrian subset, Doha and Riyadh on the Saudi subset.
- Dialect neurons take up less than 1% of MLP dimensions, yet cover only about 5–21% of the residual dialect direction. They separate MSA from the spoken dialects much more clearly than they separate dialects from one another.
- From MSA prompts, neuron steering fails completely: dialect authenticity falls to the lowest score. Vector steering still induces the dialect, and neither method adds meaningful inference cost.
- Human annotators found a trade-off the LLM judge missed: for Fanar on Moroccan, steered outputs were less fluent and coherent despite higher authenticity. Performance was poorest for the dialects that had the least data available for extracting neurons and vectors.

## Interesting ideas
- Instead of only asking whether selected units work, the authors measure how much of a distributed feature direction those units actually span; this coverage figure predicts and explains the causal limits of unit-level edits. — For Mova (learn): When we claim a few features or rules "carry" a phenomenon in the perceptron parser, we could compute what share of the relevant signal they cover before trusting edits to them.
- Testing an intervention in two conditions (the cue already present vs. absent) separates "reinforcing" a behaviour from "inducing" it, which is a sharp and cheap diagnostic. — For Mova (learn): Self-induced repair rules could be evaluated both on inputs already near the target analysis and on inputs far from it, acting as a negative control against rules that only amplify existing correct guesses.
- Units are selected by two filters at once: frequent activation for some class, and a low-entropy activation profile across classes. This avoids picking features that fire everywhere. — For Mova (tokens): The same frequent-but-selective filter could pick tagger or morphology features that truly mark a register or dialect of English, rather than generic high-frequency cues.
- The steering direction is a mean difference over parallel sentence pairs, taken from generated-response positions rather than the input, so it captures how the model produces the variety rather than how it reads it. — For Mova (back): For generation from the meaning graph, paired renderings of the same graph in two registers could yield a style offset that shifts output wording while keeping the graph fixed.
- Human annotation revealed quality losses that both an automatic dialect identifier and an LLM judge overlooked, so the authors argue for scoring fluency alongside target-attribute strength. — For Mova (learn): When LLM teacher tiers grade silver data, a small human-checked slice should track fluency or correctness separately, so that judge blind spots do not go unnoticed.

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

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{elozeiri2026can,
  title = {{Can Dialects Be Steered Like Languages? Sparse Neurons and Distributed Directions in Arabic LLMs}},
  author = {Elozeiri, Kareem and Abassy, Mervat and Kallas, Omar and Dalvi, Fahim and Nakov, Preslav and Inui, Kentaro and Durrani, Nadir},
  year = {2026},
  doi = {10.48550/arxiv.2607.03936},
  eprint = {2607.03936},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2607.03936}
}
```

- [← previous Can a Remedy Find a Researcher? Exploring the Development of Semantic Knowledge in Italian BabyLMs](https://md.research.elamur.ai/free/en/papers/suozzi-2026-can-remedy-find-researcher-exploring/)
- [next → Can Emotions Signal Gender? Investigating Implicit Cues in Human and LLM Translations of Amazon Reviews](https://md.research.elamur.ai/free/en/papers/manakhimova-2026-can-emotions-signal-gender-investigating/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/2/)

## References

26 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ahmed Abdelali, Nadir Durrani, Fahim Dalvi, and Hassan Sajjad. 2022. Post-hoc analysis of Arabic transformer models. In Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 91–103, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Ahmed+Abdelali%2C+Nadir+Durrani%2C+Fahim+Dalvi%2C+and+Hassan+Sajjad%2E+2022%2E+Post%2Dhoc+analysis+of+Arabic+transformer+models%2E+In+Proceedings+of+the+Fifth+BlackboxNLP+Workshop+on+Analyzing+and+Interpreting+Neur)
2. Muhammad Abdul-Mageed, AbdelRahim Elmadany, Chiyu Zhang, El Moatez Billah Nagoudi, Houda Bouamor, and Nizar Habash. 2023. NADI 2023: The fourth nuanced Arabic dialect identification shared task. In Proceedings of ArabicNLP 2023, pages 600–613, Singapore (Hybrid). Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Muhammad+Abdul%2DMageed%2C+AbdelRahim+Elmadany%2C+Chiyu+Zhang%2C+El+Moatez+Billah+Nagoudi%2C+Houda+Bouamor%2C+and+Nizar+Habash%2E+2023%2E+NADI+2023%3A+The+fourth+nuanced+Arabic+dialect+identification+shared+task%2E+In+Pr)
3. Krishak Aneja, Manas Mittal, Anmol Goel, Ponnurangam Kumaraguru, and Vamshi Krishna Bonagiri. 2026. Intrinsic guardrails: How semantic geometry of personality interacts with emergent misalignment in llms. Preprint, arXiv:2605.10633. [↗](https://arxiv.org/abs/2605.10633)
4. Mohamed Anwar, Abed Alhakim Freihat, George Ibrahim, Mostafa Awad, Abdelrahman Sadallah, Gurpreet Gosal, Gokulakrishnan Ramakrishnan, Sarath Chandran, Biswajit Mishra, Rituraj Joshi, Ahmed Frikha, Etienne Goffinet, Abhishek Maiti, Ali El Filali, Sarah AlBarri, Samujjwal Ghosh, Rahul Pal, Parvez Mullah, Awantika Shukla, and 41 others. 2026. Jais 2: A family of arabic-centric open large language models. Preprint, arXiv:2608.13580. [↗](https://arxiv.org/abs/2608.13580) [our card](https://md.research.elamur.ai/free/en/papers/anwar-2026-jais-2-family-arabic-centric/)
5. M Saiful Bari, Yazeed Alnumay, Norah Alzahrani, Nouf Alotaibi, Hisham Alyahya, AlRashed Al-Rashed, Faisal Mirza, Shaykhah Alsubaie, Hassan Alahmed, Ghadah Alabduljabbar, Raghad Alkhathran, Yousef Almushayqih, Raneem Alnajim, Salman I Alsubaihi, Maryam Al Mansour, Saad Hassan, Majed Alrubaian, Ali Alammari, Zaki Alawami, and 7 others. 2025. Allam: Large language models for arabic and english. In International Conference on Learning Representations, volume 2025, pages 34179–34214. [find](https://scholar.google.com/scholar?q=M+Saiful+Bari%2C+Yazeed+Alnumay%2C+Norah+Alzahrani%2C+Nouf+Alotaibi%2C+Hisham+Alyahya%2C+AlRashed+Al%2DRashed%2C+Faisal+Mirza%2C+Shaykhah+Alsubaie%2C+Hassan+Alahmed%2C+Ghadah+Alabduljabbar%2C+Raghad+Alkhathran%2C+Yousef+Almu) [our card](https://md.research.elamur.ai/free/en/papers/bari-2024-allam-large-language-models-arabic/)
6. Houda Bouamor, Nizar Habash, Mohammad Salameh, Wajdi Zaghouani, Owen Rambow, Dana Abdulrahim, Ossama Obeid, Salam Khalifa, Fadhl Eryani, Alexander Erdmann, and Kemal Oflazer. 2018. The MADAR Arabic dialect corpus and lexicon. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). [find](https://scholar.google.com/scholar?q=Houda+Bouamor%2C+Nizar+Habash%2C+Mohammad+Salameh%2C+Wajdi+Zaghouani%2C+Owen+Rambow%2C+Dana+Abdulrahim%2C+Ossama+Obeid%2C+Salam+Khalifa%2C+Fadhl+Eryani%2C+Alexander+Erdmann%2C+and+Kemal+Oflazer%2E+2018%2E+The+MADAR+Arabic+di)
7. Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. 2025. Persona vectors: Monitoring and controlling character traits in language models. arXiv preprint arXiv:2507.21509. [↗](https://arxiv.org/abs/2507.21509)
8. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, Sam Petulla, Colin Gaffney, Asaf Aharoni, Nathan Lintz, Tiago Cardal Pais, Henrik Jacobsson, Idan Szpektor, Nan-Jiang Jiang, and 3416 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Preprint, arXiv:2507.06261. [↗](https://arxiv.org/abs/2507.06261) [our card](https://md.research.elamur.ai/free/en/papers/comanici-2025-gemini-2-5-pushing-frontier/)
9. Nadir Durrani, Fahim Dalvi, and Hassan Sajjad. 2023. Discovering salient neurons in deep nlp models. Journal of Machine Learning Research, 24(362):1– 40. [find](https://scholar.google.com/scholar?q=Nadir+Durrani%2C+Fahim+Dalvi%2C+and+Hassan+Sajjad%2E+2023%2E+Discovering+salient+neurons+in+deep+nlp+models%2E+Journal+of+Machine+Learning+Research%2C+24%28362%29%3A1%E2%80%93+40%2E)
10. Nadir Durrani, Hassan Sajjad, Fahim Dalvi, and Firoj Alam. 2022. On the transformation of latent space in fine-tuned nlp models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1495–1516, Abu Dhabi, UAE. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Nadir+Durrani%2C+Hassan+Sajjad%2C+Fahim+Dalvi%2C+and+Firoj+Alam%2E+2022%2E+On+the+transformation+of+latent+space+in+fine%2Dtuned+nlp+models%2E+In+Proceedings+of+the+2022+Conference+on+Empirical+Methods+in+Natural+L)
11. Nadir Durrani, Hassan Sajjad, Fahim Dalvi, and Yonatan Belinkov. 2020. Analyzing individual neurons in pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4865–4880, Online. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Nadir+Durrani%2C+Hassan+Sajjad%2C+Fahim+Dalvi%2C+and+Yonatan+Belinkov%2E+2020%2E+Analyzing+individual+neurons+in+pre%2Dtrained+language+models%2E+In+Proceedings+of+the+2020+Conference+on+Empirical+Methods+in+Natura)
12. Mahmoud El-Haj. 2020. Habibi - a multi dialect multi national Arabic song lyrics corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 1318–1326, Marseille, France. European Language Resources Association. [find](https://scholar.google.com/scholar?q=Mahmoud+El%2DHaj%2E+2020%2E+Habibi+%2D+a+multi+dialect+multi+national+Arabic+song+lyrics+corpus%2E+In+Proceedings+of+the+Twelfth+Language+Resources+and+Evaluation+Conference%2C+pages+1318%E2%80%931326%2C+Marseille%2C+France%2E) [our card](https://md.research.elamur.ai/free/en/papers/haj-2020-habibi-multi-dialect-multi-national/)
13. Ahmed Elshabrawy, Hour Kaing, Haiyue Song, Alham Fikri Aji, Hideki Tanaka, Masao Utiyama, and Raj Dabre. 2025. When alignment hurts: Decoupling representational spaces in multilingual models. Preprint, arXiv:2508.12803. [↗](https://arxiv.org/abs/2508.12803)
14. Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad, Firoj Alam, Enes Altinisik, Ehsannedin Asgari, Yazan Boshmaf, Sabri Boughorbel, Sanjay Chawla, Shammur Chowdhury, Fahim Dalvi, Kareem Darwish, Nadir Durrani, Mohamed Elfeky, Ahmed Elmagarmid, Mohamed Eltabakh, Masoomali Fatehkia, Anastasios Fragkopoulos, Maram Hasanain, and 23 others. 2025. Fanar: An arabic-centric multimodal generative ai platform. [find](https://scholar.google.com/scholar?q=Fanar+Team%2C+Ummar+Abbas%2C+Mohammad+Shahmeer+Ahmad%2C+Firoj+Alam%2C+Enes+Altinisik%2C+Ehsannedin+Asgari%2C+Yazan+Boshmaf%2C+Sabri+Boughorbel%2C+Sanjay+Chawla%2C+Shammur+Chowdhury%2C+Fahim+Dalvi%2C+Kareem+Darwish%2C+Nadir+D) [our card](https://md.research.elamur.ai/free/en/papers/abbas-2025-fanar-arabic-centric-multimodal-generative/)
15. Daniil Gurgurov, Katharina Trinley, Yusser Al Ghussin, Tanja Baeumel, Josef Van Genabith, and Simon Ostermann. 2025. Language arithmetics: Towards systematic language neuron identification and manipulation. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, pages 2911–2937, Mumbai, India. The Asian Federation of Natural Language Processing and The Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Daniil+Gurgurov%2C+Katharina+Trinley%2C+Yusser+Al+Ghussin%2C+Tanja+Baeumel%2C+Josef+Van+Genabith%2C+and+Simon+Ostermann%2E+2025%2E+Language+arithmetics%3A+Towards+systematic+language+neuron+identification+and+manipul)
16. Nizar Y. Habash. 2010. Introduction to Arabic Natural Language Processing, volume 3 of Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers. [find](https://scholar.google.com/scholar?q=Nizar+Y%2E+Habash%2E+2010%2E+Introduction+to+Arabic+Natural+Language+Processing%2C+volume+3+of+Synthesis+Lectures+on+Human+Language+Technologies%2E+Morgan+%26+Claypool+Publishers%2E)
17. Abdulmuizz Khalak, Abderrahmane Issam, and Gerasimos Spanakis. 2026. From FusHa to folk: Exploring cross-lingual transfer in Arabic language models. In Proceedings of the 13th Workshop on NLP for Similar Languages, Varieties and Dialects, pages 196–209, Rabat, Morocco. Association for Computational Linguistics. Basel Mousi, Nadir Durrani, Fatema Ahmad, Md. [find](https://scholar.google.com/scholar?q=Abdulmuizz+Khalak%2C+Abderrahmane+Issam%2C+and+Gerasimos+Spanakis%2E+2026%2E+From+FusHa+to+folk%3A+Exploring+cross%2Dlingual+transfer+in+Arabic+language+models%2E+In+Proceedings+of+the+13th+Workshop+on+NLP+for+Simi)
18. Arid Hasan, Maram Hasanain, Tameem Kabbani, Fahim Dalvi, Shammur Absar Chowdhury, and Firoj Alam. 2025. AraDiCE: Benchmarks for dialectal and cultural capabilities in LLMs. In Proceedings of the 31st International Conference on Computational Linguistics, pages 4186–4218, Abu Dhabi, UAE. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Arid+Hasan%2C+Maram+Hasanain%2C+Tameem+Kabbani%2C+Fahim+Dalvi%2C+Shammur+Absar+Chowdhury%2C+and+Firoj+Alam%2E+2025%2E+AraDiCE%3A+Benchmarks+for+dialectal+and+cultural+capabilities+in+LLMs%2E+In+Proceedings+of+the+31st+)
19. Omer Nacar. 2025. Ui-level evaluation of allam 34b: Measuring an arabic-centric llm via humain chat. Preprint, arXiv:2508.17378. [↗](https://arxiv.org/abs/2508.17378) [our card](https://md.research.elamur.ai/free/en/papers/nacar-2025-ui-level-evaluation-allam-34b/)
20. Inaya Rahmanisa, Lyzander Marciano Andrylie, Mahardika Krisna Ihsani, Alfan Farizki Wicaksono, Haryo Akbarianto Wibowo, and Alham Fikri Aji. 2025. Unveiling the influence of amplifying language-specific neurons. In Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics, pages 919–968, Mumbai, India. The Asian Federation of Natural Language Processing and The Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Inaya+Rahmanisa%2C+Lyzander+Marciano+Andrylie%2C+Mahardika+Krisna+Ihsani%2C+Alfan+Farizki+Wicaksono%2C+Haryo+Akbarianto+Wibowo%2C+and+Alham+Fikri+Aji%2E+2025%2E+Unveiling+the+influence+of+amplifying+language%2Dspecif)
21. 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, Thailand. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Nina+Rimsky%2C+Nick+Gabrieli%2C+Julian+Schulz%2C+Meg+Tong%2C+Evan+Hubinger%2C+and+Alexander+Turner%2E+2024%2E+Steering+llama+2+via+contrastive+activation+addition%2E+In+Proceedings+of+the+62nd+Annual+Meeting+of+the+A)
22. Nathaniel Romney Robinson, Shahd Abdelmoneim, Kelly Marchisio, and Sebastian Ruder. 2025. AL-QASIDA: Analyzing LLM quality and accuracy systematically in dialectal Arabic. In Findings of the Association for Computational Linguistics: ACL 2025, pages 22048–22065, Vienna, Austria. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Nathaniel+Romney+Robinson%2C+Shahd+Abdelmoneim%2C+Kelly+Marchisio%2C+and+Sebastian+Ruder%2E+2025%2E+AL%2DQASIDA%3A+Analyzing+LLM+quality+and+accuracy+systematically+in+dialectal+Arabic%2E+In+Findings+of+the+Associati)
23. Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, and Ji-Rong Wen. 2024. Language-specific neurons: The key to multilingual capabilities in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5701–5715, Bangkok, Thailand. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Tianyi+Tang%2C+Wenyang+Luo%2C+Haoyang+Huang%2C+Dongdong+Zhang%2C+Xiaolei+Wang%2C+Xin+Zhao%2C+Furu+Wei%2C+and+Ji%2DRong+Wen%2E+2024%2E+Language%2Dspecific+neurons%3A+The+key+to+multilingual+capabilities+in+large+language+mode) [our card](https://md.research.elamur.ai/free/en/papers/tang-2024-language-specific-neurons-key-multilingual/)
24. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and 20 others. 2022. No language left behind: Scaling human-centered machine translation. Preprint, arXiv:2207.04672. [↗](https://arxiv.org/abs/2207.04672)
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. arXiv preprint arXiv:2308.10248. [↗](https://arxiv.org/abs/2308.10248) [our card](https://md.research.elamur.ai/free/en/papers/turner-2023-steering-language-models-activation-engineering/)
26. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, and 1 others. 2025. Representation engineering: A top-down approach to AI transparency. arXiv preprint arXiv:2310.01405. [↗](https://arxiv.org/abs/2310.01405) [our card](https://md.research.elamur.ai/free/en/papers/zou-2023-representation-engineering-top-down-approach/)