Elamur Research⌕
☰
home / papers / 2026 / Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions

Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions

Paper · 2026 · Ignacio Sastre and Aiala Rosá

Summary

The paper asks whether a frozen LLM can absorb a new concept from definitions alone, without labeled behavioral examples. The authors propose Concept Tokens. A single new special token is added to the vocabulary, and every mention of the target concept in a set of natural-language definitions is replaced by it. Only the token's input embedding is trained, using the ordinary language-modeling loss. Because several definitions must be predicted at once, the embedding is pushed toward a generalized representation rather than memorizing one text. Earlier memory tokens are a special case with a single definition.

Experiments use Llama 3.1 8B Instruct (4-bit) on three tasks:

  • closed-book QA on HotpotQA,
  • inducing recasting, a second-language feedback strategy,
  • a qualitative case study.

Placing the token in a prompt works directionally: negating it suppresses the associated behavior, while asserting it amplifies it. In the recasting task, the token follows other instructions better than putting the whole definitional corpus in context. The motivation is compact, privacy-friendly steering for small on-premise models, such as those used in schools.

Key points

  • On HotpotQA, negating the hallucination token reduces hallucinated answers mostly by making the model abstain more often. Asserting the token increases hallucinations and lowers precision, so the gain is in coverage trade-off rather than answer reliability.
  • In the recasting task, some prompting baselines enforce the behavior more strongly but break other instructions, such as asking follow-up questions or not correcting answers that are already right. The concept token preserves those instructions better. The dataset comes from Ceibal en Inglés.
  • The qualitative study compares a token for the real Eiffel Tower with one for a fictional "Austral Tower" in Montevideo. The fictional token captures a coherent theme but invents details such as names, dates and heights, which suggests the method steers meaning rather than storing new facts.
  • The definitional corpora were synthetically generated by stronger LLMs. Training is costly because it requires backpropagation through the whole network, and the work ran on limited compute (ClusterUY, Colab Pro). Code and data are at github.com/nsuruguay05/concept_tokens.

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

🎓Universidad de la República de UruguayMontevideo · Uruguay9 works · 450 citations

 Fields, topics and concepts

📚 Shelves 5The 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 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Ignacio Sastre and Aiala Rosá. 2026. Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1319

@inproceedings{sastre2026concept,
  title = {{Concept Tokens: Learning Behavioral Embeddings Through Concept Definitions}},
  author = {Ignacio Sastre and Aiala Rosá},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1319},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1319}
}

References

23 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L. Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, and 6 others. 2… find
  2. Romina Brown, Santiago Paez, Gonzalo Herrera, Luis Chiruzzo, and Aiala Rosá. 2023. Experiments on automatic error detection and correction for uruguayan https://ceibal.edu.uy https://cluster.uy learners of English. In Proceedings of the 12th Workshop on NLP for Computer Assisted Language Learning, p… ↗
  3. Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. 2025. Persona vectors: Monitoring and controlling character traits in language models. Preprint, arXiv:2507.21509. — [arXiv:2507.21509](https://arxiv.org/abs/2507.21509) find
  4. Ikhyun Cho and Julia Hockenmaier. 2025. Toward efficient sparse autoencoder-guided steering for improved in-context learning in large language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 28949–28961, Suzhou, China. Association for Computat… find
  5. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capa…
  6. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  7. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Trans. Inf. Syst., 43(2).
  8. Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, and Edwin Zhang. 2025. Why language models hallucinate. Preprint, arXiv:2509.04664. — [arXiv:2509.04664](https://arxiv.org/abs/2509.04664) find
  9. Yuri Kuratov, Mikhail Arkhipov, Aydar Bulatov, and Mikhail Burtsev. 2025. Cramming 1568 tokens into a single vector and back again: Exploring the limits of embedding space capacity. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa… find
  10. Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguist…
  11. Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582–… find
  12. Zongqian Li, Yinhong Liu, Yixuan Su, and Nigel Collier. 2025a. Prompt compression for large language models: A survey. 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), p…
  13. Zongqian Li, Yixuan Su, and Nigel Collier. 2025b. 500xCompressor: Generalized prompt compression for large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 25081–25091, Vienna, Austria. Association for Computat…
  14. Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61–6…
  15. Sergio Nesmachnow and Santiago Iturriaga. 2019. Cluster-uy: Collaborative scientific high performance computing in uruguay. In Supercomputing, pages 188–202, Cham. Springer International Publishing. find
  16. Howard Nicholas, Patsy Lightbown, and Nina Spada. 2002. Recasts as feedback to language learners. Language Learning, 51:719 – 758. find
  17. Ignacio Sastre and Aiala Rosá. 2025. Memory tokens: Large language models can generate reversible sentence embeddings. In Proceedings of the First Workshop on Large Language Model Memorization (L2M2), pages 183–189, Vienna, Austria. Association for Computational Linguistics. find
  18. Nishant Subramani, Nivedita Suresh, and Matthew Peters. 2022. Extracting latent steering vectors from pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566–581, Dublin, Ireland. Association for Computational Linguistics.
  19. LCM team, Loïc Barrault, Paul-Ambroise Duquenne, Maha Elbayad, Artyom Kozhevnikov, Belen Alastruey, Pierre Andrews, Mariano Coria, Guillaume Couairon, Marta R. Costa-jussà, David Dale, Hady Elsahar, Kevin Heffernan, João Maria Janeiro, Tuan Tran, Christophe Ropers, Eduardo Sánchez, Robin San Roman,… — [arXiv:2412.08821](https://arxiv.org/abs/2412.08821) find
  20. Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, and 1 others. 2024. Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet. Transformer Circuits Thread. find
  21. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369…
  22. Hanyu Zhang, Xiting Wang, Chengao Li, Xiang Ao, and Qing He. 2025. Controlling large language models through concept activation vectors. In Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence a… find
  23. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sastre-2026-concept-tokens-learning-behavioral-embeddings/