--- title: "Compositional Steering of Large Language Models with Steering" description: "Most steering research handles one behavior at a time, so this paper studies multi-behavior (compositional) steering. 43 references." authors: ["Gorjan Radevski", "Kiril Gashteovski", "Giwon Hong", "Carolin Lawrence", "Goran Glavaš"] year: 2026 venue: "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)" source: "https://aclanthology.org/2026.acl-long.1435/" doi: "10.18653/v1/2026.acl-long.1435" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/radevski-2026-compositional-steering-large-language-models/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/radevski-2026-compositional-steering-large-language-models/) [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/25/) / Compositional Steering of Large Language Models with Steering Tokens # Compositional Steering of Large Language Models with Steering Tokens **Authors:** Gorjan Radevski, [Kiril Gashteovski](https://md.research.elamur.ai/free/en/authors/kiril-gashteovski/), [Giwon Hong](https://md.research.elamur.ai/free/en/authors/giwon-hong/), [Carolin Lawrence](https://md.research.elamur.ai/free/en/authors/carolin-lawrence/) and [Goran Glavaš](https://md.research.elamur.ai/free/en/authors/goran-glavas/) · **Year:** 2026 · **Venue:** [Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)](https://md.research.elamur.ai/free/en/venues/acl/) ([ACL 2026](https://md.research.elamur.ai/free/en/venues/acl/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.acl-long.1435/) · [PDF](https://aclanthology.org/2026.acl-long.1435.pdf) · [DOI 10.18653/v1/2026.acl-long.1435](https://doi.org/10.18653/v1/2026.acl-long.1435) · [OpenAlex](https://openalex.org/W7166852544) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary Practical LLM deployments often need outputs that satisfy several requirements at once, such as a target language, length and format. Most steering research handles one behavior at a time, so this paper studies multi-behavior (compositional) steering. The authors keep the base LLM frozen and learn one input-embedding vector per behavior. Each vector is trained through self-distillation: the model's output distribution when given the explicit natural-language instruction serves as the target. In a second stage, the behavior vectors are frozen and a single extra `` token is trained on pairs of behaviors so that it learns composition as a general operation. The evaluation uses automatically verifiable constraints across several model families ranging from 3B to 14B parameters. Steering tokens beat plain instructions, activation steering and LoRA merging, especially on combinations not seen during training. Pairing the tokens with text instructions (a hybrid setup) performs best. This matters because the approach adds only a handful of trainable vectors and gives cheap, reusable control. ## Key points - **Parameter cost:** Only |B|+1 vectors of model hidden size are learned, one per behavior plus the `` token. All LLM weights stay frozen. - **Generalization:** The `` token is trained only on 2-behavior pairs. It still transfers to unseen pairings, to pairings that include unseen behaviors, and to 3-behavior compositions. - **Behaviors tested:** Output language (French, Spanish, Italian, Portuguese, German), letter case (upper, lower, title), word-count ranges and sentence counts. - **Reported gains:** For Qwen3-14B and Llama3-8B, the hybrid of tokens plus text shows gains of roughly 12–23% over text-only prompting on the strongest unseen combinations. Accuracy and robustness both increase with model size. - **Ablations:** A learned `` operator is required for compositional generalization, and orthogonality regularization is critical for zero-shot fusion. Stated limitations are that only verifiable constraints were tested, compositions went up to three behaviors at most, and models reached at most 14B parameters. *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 Edinburgh](https://md.research.elamur.ai/free/en/universities/university-of-edinburgh/) — University, Edinburgh, United Kingdom · 359 works · 77,352 citations - [University of Würzburg](https://md.research.elamur.ai/free/en/universities/university-of-wurzburg/) — University, Würzburg, Germany · 8 works · 60 citations - [Ss. Cyril and Methodius University in Skopje](https://md.research.elamur.ai/free/en/universities/ss-cyril-and-methodius-university-in-skopje/) — University, Skopje, North Macedonia · 5 works · 147 citations - [NEC (Germany)](https://md.research.elamur.ai/free/en/companies/nec-germany/) — Company, Heidelberg, Germany · 3 works · 114 citations 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works 🧭 Topics (1): [🧠 Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) · 3,320 works ## Cite Cite as: Gorjan Radevski, Kiril Gashteovski, Giwon Hong, Carolin Lawrence and Goran Glavaš. 2026. Compositional Steering of Large Language Models with Steering Tokens. Proceedings of the 64th Annual Meeting 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. ```bibtex @inproceedings{radevski2026compositional, title = {{Compositional Steering of Large Language Models with Steering Tokens}}, author = {Gorjan Radevski, Kiril Gashteovski, Giwon Hong, Carolin Lawrence and Goran Glavaš}, booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, year = {2026}, doi = {10.18653/v1/2026.acl-long.1435}, url = {https://doi.org/10.18653/v1/2026.acl-long.1435} } ``` - [← previous Compositional Meaning Representations in LLMs: a Critical Review of Probing Studies](https://md.research.elamur.ai/free/en/papers/marro-2026-compositional-meaning-representations-llms-critical/) - [next → Comprehensive Benchmarking of Long-Form Speech Generation in Diverse Scenarios](https://md.research.elamur.ai/free/en/papers/pan-2026-comprehensive-benchmarking-long-form-speech/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/25/) ## References 43 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martín Blázquez, Guilherme Penedo, Lewis Tunstall, Andrés Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, and 1 others. 2025. [SmolLM2: When Smol Goes Big– Data-Centric Training of a Small Language Model](https://md.research.elamur.ai/free/en/papers/allal-2025-smollm2-smol-goes-big-data/). In Conference on… 2. Alan Ansell, Ivan Vulić, Hannah Sterz, Anna Korhonen, and Edoardo M Ponti. 2024. Scaling Sparse Fine-Tuning to Large Language Models. arXiv preprint arXiv:2401.16405. — [arXiv:2401.16405](https://arxiv.org/abs/2401.16405) [find](https://scholar.google.com/scholar?q=Alan+Ansell%2C+Ivan+Vulic%CC%81%2C+Hannah+Sterz%2C+Anna+Korhonen%2C+and+Edoardo+M+Ponti%2E+2024%2E+Scaling+Sparse+Fine%2DTuning+to+Large+Language+Models%2E+arXiv+preprint+arXiv%3A2401%2E16405%2E+%E2%80%94+%5BarXiv%3A2401%2E16405%5D%28https%3A%2F%2Farx) 3. Adaptation of Large Language Models. In International Conference on Learning Representations (ICLR). [find](https://scholar.google.com/scholar?q=Adaptation+of+Large+Language+Models%2E+In+International+Conference+on+Learning+Representations+%28ICLR%29%2E) 4. Shawn Im and Yixuan Li. 2025. A Unified Understanding and Evaluation of Steering Methods. arXiv preprint arXiv:2502.02716. — [arXiv:2502.02716](https://arxiv.org/abs/2502.02716) [find](https://scholar.google.com/scholar?q=Shawn+Im+and+Yixuan+Li%2E+2025%2E+A+Unified+Understanding+and+Evaluation+of+Steering+Methods%2E+arXiv+preprint+arXiv%3A2502%2E02716%2E+%E2%80%94+%5BarXiv%3A2502%2E02716%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2502%2E02716%29) 5. Daniel Jaroslawicz, Brendan Whiting, Parth Shah, and Karime Maamari. 2025. How Many Instructions Can LLMs Follow at Once? In Workshop: Evaluating the Evolving LLM Lifecycle: Benchmarks, Emergent Abilities, and Scaling@NeurIPS. [find](https://scholar.google.com/scholar?q=Daniel+Jaroslawicz%2C+Brendan+Whiting%2C+Parth+Shah%2C+and+Karime+Maamari%2E+2025%2E+How+Many+Instructions+Can+LLMs+Follow+at+Once%3F+In+Workshop%3A+Evaluating+the+Evolving+LLM+Lifecycle%3A+Benchmarks%2C+Emergent+Abili) 6. Reza Bayat, Ali Rahimi-Kalahroudi, Mohammad Pezeshki, Sarath Chandar, and Pascal Vincent. 2025. Steering Large Language Model Activations in Sparse Spaces. In Conference on Language Modeling (COLM). [find](https://scholar.google.com/scholar?q=Reza+Bayat%2C+Ali+Rahimi%2DKalahroudi%2C+Mohammad+Pezeshki%2C+Sarath+Chandar%2C+and+Pascal+Vincent%2E+2025%2E+Steering+Large+Language+Model+Activations+in+Sparse+Spaces%2E+In+Conference+on+Language+Modeling+%28COLM%29%2E) 7. Yichen Jiang, Marco Vecchio, Mohit Bansal, and Anders Johannsen. 2024. Hierarchical and Dynamic Prompt Compression for Efficient Zero-Shot API Usage. In Findings of the Association for Computational Linguistics: EACL 2024, pages 2162–2174. [find](https://scholar.google.com/scholar?q=Yichen+Jiang%2C+Marco+Vecchio%2C+Mohit+Bansal%2C+and+Anders+Johannsen%2E+2024%2E+Hierarchical+and+Dynamic+Prompt+Compression+for+Efficient+Zero%2DShot+API+Usage%2E+In+Findings+of+the+Association+for+Computational+L) 8. Yuanpu Cao, Tianrong Zhang, Bochuan Cao, Ziyi Yin, Lu Lin, Fenglong Ma, and Jinghui Chen. 2024. Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-Directional Preference Optimization. Advances in Neural Information Processing Systems (NeurIPS), 37:49519–49551. [find](https://scholar.google.com/scholar?q=Yuanpu+Cao%2C+Tianrong+Zhang%2C+Bochuan+Cao%2C+Ziyi+Yin%2C+Lu+Lin%2C+Fenglong+Ma%2C+and+Jinghui+Chen%2E+2024%2E+Personalized+Steering+of+Large+Language+Models%3A+Versatile+Steering+Vectors+Through+Bi%2DDirectional+Prefer) 9. Mehdi Khamassi, Marceau Nahon, and Raja Chatila. 2024. Strong and Weak Alignment of Large Language Models with Human Values. Scientific Reports, 14(1):19399. [find](https://scholar.google.com/scholar?q=Mehdi+Khamassi%2C+Marceau+Nahon%2C+and+Raja+Chatila%2E+2024%2E+Strong+and+Weak+Alignment+of+Large+Language+Models+with+Human+Values%2E+Scientific+Reports%2C+14%281%29%3A19399%2E) 10. Patrick Queiroz Da Silva, Hari Sethuraman, Dheeraj Rajagopal, Hannaneh Hajishirzi, and Sachin Kumar. 2025. Steering off Course: Reliability Challenges in Steering Language Models. In Association for Computational Linguistics (ACL), page 19856–19882. [find](https://scholar.google.com/scholar?q=Patrick+Queiroz+Da+Silva%2C+Hari+Sethuraman%2C+Dheeraj+Rajagopal%2C+Hannaneh+Hajishirzi%2C+and+Sachin+Kumar%2E+2025%2E+Steering+off+Course%3A+Reliability+Challenges+in+Steering+Language+Models%2E+In+Association+for+C) 11. Jiancheng Dong, Pengyue Jia, Jingyu Peng, Maolin Wang, Yuhao Wang, Lixin Su, Xin Sun, Shuaiqiang Wang, Dawei Yin, and Xiangyu Zhao. 2025. Behavior-Equivalent Token: Single-Token Replacement for Long Prompts in LLMs. arXiv preprint arXiv:2511.23271. — [arXiv:2511.23271](https://arxiv.org/abs/2511.23271) [find](https://scholar.google.com/scholar?q=Jiancheng+Dong%2C+Pengyue+Jia%2C+Jingyu+Peng%2C+Maolin+Wang%2C+Yuhao+Wang%2C+Lixin+Su%2C+Xin+Sun%2C+Shuaiqiang+Wang%2C+Dawei+Yin%2C+and+Xiangyu+Zhao%2E+2025%2E+Behavior%2DEquivalent+Token%3A+Single%2DToken+Replacement+for+Long+P) 12. Federico Errica, Davide Sanvito, Giuseppe Siracusano, and Roberto Bifulco. 2025. [What Did i Do Wrong?](https://md.research.elamur.ai/free/en/papers/errica-2025-what-did-i-do-wrong/) Quantifying LLMs’ Sensitivity and Consistency to Prompt Engineering. In Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pag… 13. Shivanshu Gupta, Clemens Rosenbaum, and Ethan R Elenberg. 2024. GistScore: Learning Better Representations for In-Context Example Selection with Gist Bottlenecks. In Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo@ICLR). [find](https://scholar.google.com/scholar?q=Shivanshu+Gupta%2C+Clemens+Rosenbaum%2C+and+Ethan+R+Elenberg%2E+2024%2E+GistScore%3A+Learning+Better+Representations+for+In%2DContext+Example+Selection+with+Gist+Bottlenecks%2E+In+Workshop+on+Mathematical+and+Empir) 14. Chi Han, Jialiang Xu, Manling Li, Yi Fung, Chenkai Sun, Nan Jiang, Tarek Abdelzaher, and Heng Ji. 2024. [Word embeddings are steers for language models](https://md.research.elamur.ai/free/en/papers/han-2024-word-embeddings-are-steers-language/). In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16410–16430. 15. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. [find](https://scholar.google.com/scholar?q=Edward+J+Hu%2C+Yelong+Shen%2C+Phillip+Wallis%2C+Zeyuan+Allen%2DZhu%2C+Yuanzhi+Li%2C+Shean+Wang%2C+Lu+Wang%2C+Weizhu+Chen%2C+and+1+others%2E+2022%2E) 16. LoRA: Low-Rank Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. [The power of scale for parameter-efficient prompt tuning](https://md.research.elamur.ai/free/en/papers/lester-2021-power-scale-parameter-efficient-prompt/). In Proceedings of the 2021 conference on empirical methods in natural language processing, pages 3045–3059. 17. Xiang Lisa Li and Percy Liang. 2021. Prefix-Tuning: Optimizing Continuous Prompts for Generation. arXiv preprint arXiv:2101.00190. [find](https://scholar.google.com/scholar?q=Xiang+Lisa+Li+and+Percy+Liang%2E+2021%2E+Prefix%2DTuning%3A+Optimizing+Continuous+Prompts+for+Generation%2E+arXiv+preprint+arXiv%3A2101%2E00190%2E) 18. Xinze Li, Zhenghao Liu, Chenyan Xiong, Shi Yu, Yukun Yan, Shuo Wang, and Ge Yu. 2024. Say More with Less: Understanding Prompt Learning Behaviors through Gist Compression. arXiv preprint arXiv:2402.16058. — [arXiv:2402.16058](https://arxiv.org/abs/2402.16058) [find](https://scholar.google.com/scholar?q=Xinze+Li%2C+Zhenghao+Liu%2C+Chenyan+Xiong%2C+Shi+Yu%2C+Yukun+Yan%2C+Shuo+Wang%2C+and+Ge+Yu%2E+2024%2E+Say+More+with+Less%3A+Understanding+Prompt+Learning+Behaviors+through+Gist+Compression%2E+arXiv+preprint+arXiv%3A2402%2E16) 19. Zongqian Li, Yinhong Liu, Yixuan Su, and Nigel Collier. 2025. Prompt Compression for [Large Language Models: A Survey](https://md.research.elamur.ai/free/en/papers/minaee-2024-large-language-models-survey/). In Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 7182–7195. 20. Seungwon Lim, Seungbeen Lee, Dongjun Min, and Youngjae Yu. 2025. Persona Dynamics: Unveiling the Impact of Personality Traits on Agents in Text-Based Games. In Association for Computational Linguistics (ACL), page 31360–31394. [find](https://scholar.google.com/scholar?q=Seungwon+Lim%2C+Seungbeen+Lee%2C+Dongjun+Min%2C+and+Youngjae+Yu%2E+2025%2E+Persona+Dynamics%3A+Unveiling+the+Impact+of+Personality+Traits+on+Agents+in+Text%2DBased+Games%2E+In+Association+for+Computational+Linguistic) 21. Arash Marioriyad, Mohammad Hossein Rohban, and Mahdieh Soleymani Baghshah. 2025. The Silent Judge: Unacknowledged Shortcut Bias in LLM-as-a-Judge. In Workshop on Reliable ML from Unreliable Data@NeurIPS. [find](https://scholar.google.com/scholar?q=Arash+Marioriyad%2C+Mohammad+Hossein+Rohban%2C+and+Mahdieh+Soleymani+Baghshah%2E+2025%2E+The+Silent+Judge%3A+Unacknowledged+Shortcut+Bias+in+LLM%2Das%2Da%2DJudge%2E+In+Workshop+on+Reliable+ML+from+Unreliable+Data%40NeurI) 22. Jesse Mu, Xiang Li, and Noah Goodman. 2023. [Learning to Compress Prompts with Gist Tokens](https://md.research.elamur.ai/free/en/papers/mu-2023-learning-compress-prompts-gist-tokens/). Advances in Neural Information Processing Systems (NeurIPS), 36:19327–19352. 23. Duy Nguyen, Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. 2025. [Multi-Attribute Steering of Language Models via Targeted Intervention](https://md.research.elamur.ai/free/en/papers/nguyen-2025-multi-attribute-steering-language-models/). In Association for Computational Linguistics (ACL), page 20619–20634. 24. Lilian Ngweta, Kiran Kate, Jason Tsay, and Yara Rizk. 2025. Towards LLMs Robustness to Changes in Prompt Format Styles. In Student Research Workshop @ Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 529–537. [find](https://scholar.google.com/scholar?q=Lilian+Ngweta%2C+Kiran+Kate%2C+Jason+Tsay%2C+and+Yara+Rizk%2E+2025%2E+Towards+LLMs+Robustness+to+Changes+in+Prompt+Format+Styles%2E+In+Student+Research+Workshop+%40+Nations+of+the+Americas+Chapter+of+the+Associatio) 25. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. [Training Language Models to Follow Instructions with Human Feedback](https://md.research.elamur.ai/free/en/papers/ouyang-2022-training-language-models-follow-instructions/). Neural Information Processing Systems (NeurIPS), 35:27730– 27744. 26. Tsung-Min Pai, Jui-I Wang, Li-Chun Lu, Shao-Hua Sun, Hung-Yi Lee, and Kai-Wei Chang. 2025. BILLY: Steering Large Language Models via Merging Persona Vectors. arXiv preprint arXiv:2510.10157. — [arXiv:2510.10157](https://arxiv.org/abs/2510.10157) [find](https://scholar.google.com/scholar?q=Tsung%2DMin+Pai%2C+Jui%2DI+Wang%2C+Li%2DChun+Lu%2C+Shao%2DHua+Sun%2C+Hung%2DYi+Lee%2C+and+Kai%2DWei+Chang%2E+2025%2E+BILLY%3A+Steering+Large+Language+Models+via+Merging+Persona+Vectors%2E+arXiv+preprint+arXiv%3A2510%2E10157%2E+%E2%80%94+%5BarXiv%3A) 27. Aleksandar Petrov, Mark Sandler, Andrey Zhmoginov, Nolan Miller, and Max Vladymyrov. 2025. Long Context In-Context Compression by Getting to the Gist of Gisting. arXiv preprint arXiv:2504.08934. — [arXiv:2504.08934](https://arxiv.org/abs/2504.08934) [find](https://scholar.google.com/scholar?q=Aleksandar+Petrov%2C+Mark+Sandler%2C+Andrey+Zhmoginov%2C+Nolan+Miller%2C+and+Max+Vladymyrov%2E+2025%2E+Long+Context+In%2DContext+Compression+by+Getting+to+the+Gist+of+Gisting%2E+arXiv+preprint+arXiv%3A2504%2E08934%2E+%E2%80%94+%5Bar) 28. Jason Phang. 2024. Investigating the Effectiveness of Hypertuning via Gisting. arXiv preprint arXiv:2402.16817. — [arXiv:2402.16817](https://arxiv.org/abs/2402.16817) [find](https://scholar.google.com/scholar?q=Jason+Phang%2E+2024%2E+Investigating+the+Effectiveness+of+Hypertuning+via+Gisting%2E+arXiv+preprint+arXiv%3A2402%2E16817%2E+%E2%80%94+%5BarXiv%3A2402%2E16817%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F2402%2E16817%29) 29. Gorjan Radevski, Dusan Grujicic, Matthew Blaschko, Marie-Francine Moens, and Tinne Tuytelaars. 2023. Multimodal Distillation for Egocentric Action Recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5213–5224. [find](https://scholar.google.com/scholar?q=Gorjan+Radevski%2C+Dusan+Grujicic%2C+Matthew+Blaschko%2C+Marie%2DFrancine+Moens%2C+and+Tinne+Tuytelaars%2E+2023%2E+Multimodal+Distillation+for+Egocentric+Action+Recognition%2E+In+Proceedings+of+the+IEEE%2FCVF+Internati) 30. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steering Llama 2 via Contrastive Activation Addition. In Association for Computational Linguistics (ACL), pages 15504–15522. [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+Association+for+Computational+Linguistics+%28ACL%29) 31. Daniel Scalena, Gabriele Sarti, and Malvina Nissim. 2024. Multi-Property Steering of Large Language Models with Dynamic Activation Composition. In BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP @ ACL, pages 577–603. [find](https://scholar.google.com/scholar?q=Daniel+Scalena%2C+Gabriele+Sarti%2C+and+Malvina+Nissim%2E+2024%2E+Multi%2DProperty+Steering+of+Large+Language+Models+with+Dynamic+Activation+Composition%2E+In+BlackboxNLP+Workshop%3A+Analyzing+and+Interpreting+Neur) 32. Zara Siddique, Irtaza Khalid, Liam D Turner, and Luis Espinosa-Anke. 2025. Shifting Perspectives: Steering Vector Ensembles for Robust Bias Mitigation in LLMs. arXiv preprint arXiv:2503.05371. — [arXiv:2503.05371](https://arxiv.org/abs/2503.05371) [find](https://scholar.google.com/scholar?q=Zara+Siddique%2C+Irtaza+Khalid%2C+Liam+D+Turner%2C+and+Luis+Espinosa%2DAnke%2E+2025%2E+Shifting+Perspectives%3A+Steering+Vector+Ensembles+for+Robust+Bias+Mitigation+in+LLMs%2E+arXiv+preprint+arXiv%3A2503%2E05371%2E+%E2%80%94+%5BarXi) 33. Hannah Sterz, Fabian David Schmidt, Goran Glavaš, and Ivan Vulić. 2025. ReCoVeR the Target Language: Language Steering without Sacrificing Task Performance. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 19390– 19405. [find](https://scholar.google.com/scholar?q=Hannah+Sterz%2C+Fabian+David+Schmidt%2C+Goran+Glava%C5%A1%2C+and+Ivan+Vulic%CC%81%2E+2025%2E+ReCoVeR+the+Target+Language%3A+Language+Steering+without+Sacrificing+Task+Performance%2E+In+Findings+of+the+Association+for+Computa) 34. Alessandro Stolfo, Vidhisha Balachandran, Safoora Yousefi, Eric Horvitz, and Besmira Nushi. 2025. Improving Instruction-Following in Language Models through Activation Steering. In International Conference on Learning Representations (ICLR). [find](https://scholar.google.com/scholar?q=Alessandro+Stolfo%2C+Vidhisha+Balachandran%2C+Safoora+Yousefi%2C+Eric+Horvitz%2C+and+Besmira+Nushi%2E+2025%2E+Improving+Instruction%2DFollowing+in+Language+Models+through+Activation+Steering%2E+In+International+Confe) 35. Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, and Meng Jiang. 2024. [Democratizing Large Language Models via Personalized Parameter-Efficient Fine-Tuning](https://md.research.elamur.ai/free/en/papers/tan-2024-democratizing-large-language-models-personalized/). In Empirical Methods in Natural Language Processing (EMNLP), pages 6476– 6491. 36. Dmitrii Tarasov, Elizaveta Goncharova, and Kuznetsov Andrey. 2025. Sentence-Anchored Gist Compression for Long-Context LLMs. arXiv preprint arXiv:2511.08128. — [arXiv:2511.08128](https://arxiv.org/abs/2511.08128) [find](https://scholar.google.com/scholar?q=Dmitrii+Tarasov%2C+Elizaveta+Goncharova%2C+and+Kuznetsov+Andrey%2E+2025%2E+Sentence%2DAnchored+Gist+Compression+for+Long%2DContext+LLMs%2E+arXiv+preprint+arXiv%3A2511%2E08128%2E+%E2%80%94+%5BarXiv%3A2511%2E08128%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs) 37. Katherine Tian, Eric Mitchell, Huaxiu Yao, Christopher D Manning, and Chelsea Finn. 2024. [Fine-Tuning Language Models for Factuality](https://md.research.elamur.ai/free/en/papers/tian-2023-fine-tuning-language-models-factuality/). In International Conference on Learning Representations (ICLR). 38. Teun van der Weij, Massimo Poesio, and Nandi Schoots. 2024. Extending Activation Steering to Broad Skills and Multiple Behaviours. arXiv preprint arXiv:2403.05767. — [arXiv:2403.05767](https://arxiv.org/abs/2403.05767) [find](https://scholar.google.com/scholar?q=Teun+van+der+Weij%2C+Massimo+Poesio%2C+and+Nandi+Schoots%2E+2024%2E+Extending+Activation+Steering+to+Broad+Skills+and+Multiple+Behaviours%2E+arXiv+preprint+arXiv%3A2403%2E05767%2E+%E2%80%94+%5BarXiv%3A2403%2E05767%5D%28https%3A%2F%2Farxiv%2Eo) 39. Zhengxuan Wu, Qinan Yu, Aryaman Arora, Christopher D Manning, and Christopher Potts. 2025. Improved representation steering for language models. arXiv preprint arXiv:2505.20809. [find](https://scholar.google.com/scholar?q=Zhengxuan+Wu%2C+Qinan+Yu%2C+Aryaman+Arora%2C+Christopher+D+Manning%2C+and+Christopher+Potts%2E+2025%2E+Improved+representation+steering+for+language+models%2E+arXiv+preprint+arXiv%3A2505%2E20809%2E) 40. Zhouhang Xie, Junda Wu, Yiran Shen, Yu Xia, Xintong Li, Aaron Chang, Ryan Rossi, Sachin Kumar, Bodhisattwa Prasad Majumder, Jingbo Shang, and 1 others. 2025. A Survey on Personalized and Pluralistic Preference Alignment in Large Language Models. In Conference on Language Modeling (COLM). [find](https://scholar.google.com/scholar?q=Zhouhang+Xie%2C+Junda+Wu%2C+Yiran+Shen%2C+Yu+Xia%2C+Xintong+Li%2C+Aaron+Chang%2C+Ryan+Rossi%2C+Sachin+Kumar%2C+Bodhisattwa+Prasad+Majumder%2C+Jingbo+Shang%2C+and+1+others%2E+2025%2E+A+Survey+on+Personalized+and+Pluralistic+P) 41. Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. [Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch](https://md.research.elamur.ai/free/en/papers/yu-2023-language-models-super-mario-absorbing/). In International Conference on Machine Learning (ICML). 42. Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Guoyin Wang, and Fei Wu. 2025. [Instruction Tuning for Large Language Models: A Survey](https://md.research.elamur.ai/free/en/papers/zhang-2023-instruction-tuning-large-language-models/). ACM Computing Surveys. 43. Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, and 1 others. 2024. [A Roadmap to Pluralistic Alignment](https://md.research.elamur.ai/free/en/papers/sorensen-2024-roadmap-pluralistic-alignment/). arXiv preprint arXiv:2402.05070. *Updated 1 November 2026*