Elamur Research⌕
☰
home / papers / 2026 / Are Formal and Functional Linguistic Mechanisms Dissociated in Language Models?

Are Formal and Functional Linguistic Mechanisms Dissociated in Language Models?

Paper · 2026 · Michael Hanna, Yonatan Belinkov and Sandro Pezzelle

Summary

This study asks whether today's LLMs internally separate formal linguistic competence (syntax, morphology, lexical semantics) from functional competence (reasoning, world knowledge, tracking situations). Neuroscience suggests the human brain has a dedicated language network distinct from regions used for thought. The authors test this idea in models using mechanistic interpretability. They define 5 formal and 5 functional tasks and identify the circuit for each in 5 LLMs. A circuit is a minimal causal subgraph that preserves model behavior when everything else is ablated. They then compare circuits in two ways. Edge intersection-over-union measures structural overlap. Cross-task faithfulness measures how well one task's circuit performs another task.

Overlap is low between formal and functional circuits, but it is also low among the formal circuits themselves. This means no single unified language circuit appears. Cross-task faithfulness, however, does split the tasks into a formal and a functional group. This suggests a shared pool of formal mechanisms from which individual tasks draw different subsets. The work matters for designing modular architectures and for judging whether brain–LLM comparisons are justified.

Key points

  • The setup covers 5 LLMs and 10 tasks. Examples include Subject-Verb Agreement (formal), and Greater-Than, IOI, Colored Objects, Entity Tracking, and Capital-Country / Country-Capital (functional). Hypernymy also appears among the tasks.
  • Circuits had to reach an 85% faithfulness threshold. A rerun at 90% gave the same patterns. In that rerun, the cross-task faithfulness dendrogram perfectly recovered the formal/functional split, while the IoU dendrogram swapped Hypernymy and Greater-Than.
  • IoU values are compressed, so small values can still be meaningful:
    • Capital-Country vs. Country-Capital reached IoU 0.5 with near-perfect cross-task faithfulness.
    • Greater-Than variants on Llama-3 (8B) reached only 0.25–0.31 IoU.
  • Control tasks (Math / Echo) overlapped with formal-task circuits at an average IoU of about 0.09. The authors take this as evidence that the formal–functional overlap is not simply because functional tasks contain formal ones.
  • Code and the circuit-finding tools are released at github.com/hannamw/formal-functional-dissociation.

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 AmsterdamAmsterdam · Netherlands144 works · 173,906 citations🎓Technion – Israel Institute of TechnologyHaifa · Israel49 works · 5,306 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksNeuroscience · 165 worksCognitive Neuroscience · 164 works

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

Cite

Cite as Michael Hanna, Yonatan Belinkov and Sandro Pezzelle. 2026. Are Formal and Functional Linguistic Mechanisms Dissociated in Language Models?. Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.1162/coli.a.24

@article{hanna2026are,
  title = {{Are Formal and Functional Linguistic Mechanisms Dissociated in Language Models?}},
  author = {Michael Hanna, Yonatan Belinkov and Sandro Pezzelle},
  journal = {Computational Linguistics},
  year = {2026},
  doi = {10.1162/coli.a.24},
  url = {https://doi.org/10.1162/coli.a.24}
}

References

26 references, 12 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmadian, Arash, Saurabh Dash, Hongyu Chen, Bharat Venkitesh, Zhen Stephen Gou, Phil Blunsom, Ahmet Üstün, and Sara Hooker. 2023. Intriguing properties of quantization at scale. In Advances in Neural Information Processing Systems, volume 36, pages 34278–34294. find
  2. AlKhamissi, Badr, Greta Tuckute, Antoine Bosselut, and Martin Schrimpf. 2024. The LLM language network: A neuroscientific approach for identifying causally task-relevant units. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguisti… find
  3. Metaphysics Research Lab, Stanford University. BigBench Team. 2023. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research. https://openreview.net/forum?id=uyTL5Bvosj Bozic, Mirjana, Lorraine K. Tyler, David T. Ives, B… ↗
  4. Marslen-Wilson. 2010. Bihemispheric foundations for human speech comprehension. Proceedings of the National Academy of Sciences, 107(40):17439–17444. https://doi.org/10.1073/pnas .1000531107, PubMed: 20855587 Chan, Lawrence, Adrià Garriga-Alonso, Nicholas Goldwosky-Dill, Ryan Greenblatt, Jenny Niti… — [doi:10.1073/pnas](https://doi.org/10.1073/pnas) ↗
  5. AI Alignment Forum. Chang, Tyler A. and Benjamin K. Bergen. 2024. Language model behavior: A comprehensive survey. Computational Linguistics, 50(1):293–350. https://doi.org/10.1162/coli a 00492 Chen, Xuanyi, Josef Affourtit, Rachel Ryskin, Tamar I. Regev, Samuel Norman-Haignere, Olessia Jouravlev,… ↗
  6. Elhage, Nelson, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. 2022. Toy models of superposition. Transforme… find
  7. Elhage, Nelson, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Deep Ganguli, Zac Hatfield-Dodds, Danny Hernandez, Andy Jones, Jackson Kernion, Liane Lovitt, Kamal Ndousse, Dario Amodei, Tom Brown, J… find
  8. Fedorenko, Evelina, Idan Asher Blank, Matthew Siegelman, and Zachary Mineroff. 2020. Lack of selectivity for syntax relative to word meanings throughout the language network. Cognition, 203:104348. https://doi.org/10.1016/j.cognition.2020.104348, PubMed: 32569894 Fedorenko, Evelina, Po-Jang Hsieh,… — [doi:10.1016/j.cognition.2020.104348](https://doi.org/10.1016/j.cognition.2020.104348) ↗
  9. Whitfield-Gabrieli, and Nancy G. Kanwisher. 2010. New method for fMRI investigations of language: Defining ROIs functionally in individual subjects. Journal of Neurophysiology, 104(2):1177–1194. https://doi.org/10.1152/jn.00032.2010, PubMed: 20410363 Fedorenko, Evelina, Anna A. Ivanova, and Tamar I… ↗
  10. Hanna, Michael, Ollie Liu, and Alexandre Variengien. 2023. How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model. In Thirty-seventh Conference on Neural Information Processing Systems. https://openreview.net/forum?id=p4PckNQR8k Hanna, Michael, San… ↗
  11. Hu, Jennifer, Jon Gauthier, Peng Qian, Ethan Wilcox, and Roger Levy. 2020. A systematic assessment of syntactic generalization in neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1725–1744. https://doi.org/10 .18653/v1/2020.acl… find
  12. Ivanova, Anna A., Shashank Srikant, Yotaro Sueoka, Hope H. Kean, Riva Dhamala, Una-May O’Reilly, Marina U. Bers, and Evelina Fedorenko. 2020. Comprehension of computer code relies primarily on domain-general executive brain regions. eLife, 9:e58906. https://doi.org/10 .7554/eLife.58906, PubMed: 3331… find
  13. Alignment Forum. Leone, Carmela, Peter Feys, Lousin Moumdjian, Emanuele D’Amico, Mario Zappia, and Francesco Patti. 2017. Cognitive-motor dual-task interference: A systematic review of neural correlates. Neuroscience & Biobehavioral Reviews, 75:348–360. https://doi.org/10.1016 /j.neubiorev.2017.01.0… find
  14. Li, Maximilian and Lucas Janson. 2024. Optimal ablation for interpretability. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. https://openreview.net/forum?id=opt72TYzwZ Lieberum, Tom, Matthew Rahtz, János Kramár, Neel Nanda, Geoffrey Irving, Rohin Shah, and Vladim… ↗
  15. Lin, Zehui, Liwei Wu, Mingxuan Wang, and Lei Li. 2021. Learning language specific sub-network for multilingual machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (V…
  16. Llama Team. 2024. The Llama 3 herd of models. arXiv. https://arxiv.org/abs/2407.21783 Mahowald, Kyle, Anna A. Ivanova, Idan A. Blank, Nancy Kanwisher, Joshua B. Tenenbaum, and Evelina Fedorenko. 2024. Dissociating language and thought in large language models. Trends in Cognitive Sciences, 28(6):51… ↗
  17. Marks, Samuel, Can Rager, Eric J. Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. 2024. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. arXiv preprint arXiv:2403.19647. https://arxiv.org/abs/2403.19647 Mason, P. H., J. F. Domı́nguez D., B. W… ↗
  18. Monti, Martin M., Lawrence M. Parsons, and Daniel N. Osherson. 2009. The boundaries of language and thought in deductive inference. Proceedings of the National Academy of Sciences, 106(30):12554–12559. https://doi.org/10.1073/pnas.0902422106, PubMed: 19617569 Murphy, Elliot and Oscar Woolnough. 202… ↗
  19. Olah, Chris, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. 2020. Zoom in: An introduction to circuits. Distill. https://doi.org/10.23915/distill.00024.001 Olshausen, Bruno A. and David J. Field. 1997. Sparse coding with an overcomplete basis set: A strategy employed… ↗
  20. Pearl, Judea. 2009. Causality: Models, Reasoning and Inference, 2nd edition. Cambridge University Press. https://doi.org/10.1017/CBO9780511803161 Prakash, Nikhil, Tamar Rott Shaham, Tal Haklay, Yonatan Belinkov, and David Bau. 2024. Fine-tuning enhances existing mechanisms: A case study on entity t… ↗
  21. Randy L. Bruckner, Lauren M. DiNicola. 2019. The brain’s default network: Updated anatomy, physiology and evolving insights. Nature Reviews Neuroscience, 20:593–608. https://doi.org/10.1038/s41583-019-0212-7, PubMed: 31492945 Regev, Mor, Christopher J. Honey, Erez Simony, and Uri Hasson. 2013. Sele… ↗
  22. Tuckute, Greta, Nancy Kanwisher, and Evelina Fedorenko. 2024. Language in brains, minds, and machines. Annual Review of Neuroscience, 47:277–301. https://doi.org/10.1146/annurev -neuro-120623-101142, PubMed: 38669478 — [doi:10.1146/annurev](https://doi.org/10.1146/annurev) ↗
  23. Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.
  24. Vig, Jesse and Yonatan Belinkov. 2019. Analyzing the structure of attention in a transformer language model. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 63–76. https://doi.org/10.18653/v1/W19-4808 Vig, Jesse, Sebastian Gehrmann, Yon… ↗
  25. Virtanen, Pauli, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al. 2020. SciPy 1.0: Fundamental algorithms for scientific computing in Python. Nature Methods, 17:261–272. https://doi.org /10.103… ↗
  26. Xu, Ziwei, Sanjay Jain, and Mohan Kankanhalli. 2024. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817. https://arxiv.org/abs/2401.11817 Yang, An, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayihen… ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/hanna-2026-are-formal-functional-linguistic-mechanisms/