Elamur Research⌕
☰
home / papers / 2025 / Sparks of Pure Competence in LLMs: the Case of Syntactic Center Embedding in English

Sparks of Pure Competence in LLMs: the Case of Syntactic Center Embedding in English

Paper · 2025 · Daniel Hardt

Summary

Linguistic theory treats center embedding as a key example of the gap between competence and performance. The grammar allows clauses nested inside clauses without limit, but people usually cannot understand more than one level because of memory limits. Hardt asks whether a strong LLM, which may not share those limits, behaves more like the idealized competence grammar. He builds synthetic English sentences from a fixed template, with relative clauses nested from level 1 to level 4. Four LLMs must answer questions about who did what to whom inside these structures. The tests vary which embedding level is queried, how many few-shot examples are given and how long the noun phrases are. A grammaticality judgment task is also included. GPT-4 stays highly accurate even at three or four levels in some settings, unlike the weaker models and unlike humans. Its accuracy still drops with zero-shot prompting, at level 4, and with some question and NP choices. The author concludes that GPT-4 is not yet a clean case of pure competence but is much closer to it than people are. This suggests LLMs could be a low-friction setting for studying grammar.

Key points

  • The data are template-generated English sentences of the form "The N the N … TV … IV":
    • nouns: teacher, student, driver, girl, man, woman, boy
    • transitive verbs: saw, hit, likes, hates, knows
    • intransitive endings: is happy, left, is glad
  • There are four tests plus a grammaticality judgment task. They vary:
    • the queried level (Q0 targets the deepest predication, Q1 another level)
    • the number of few-shot examples (e.g. P5, P25; zero-shot is also tested)
    • the length of the NPs
  • GPT-4's best settings are Q0 with P5 and Q1 with P25, where accuracy is high but not perfect and falls noticeably at level 4. Zero-shot accuracy is low.
  • A small manual error analysis of the first 10 GPT-4 items in tests 1 and 2 sorts mistakes into error types:
    • Q0 errors are mostly of one type, possibly because the subject sits right next to the queried verb.
    • Q1 errors are more mixed.
  • Limitations: only English, only one kind of center embedding (relative clauses modifying NPs), only four models, and a small informal human survey rather than rigorous human experiments.

Interesting ideas

  • I1. Probe deep structure with a template generator whose gold analysis is known in advance, then ask about the innermost predication ("who hit who") so that surface heuristics cannot pass the test.
  • I2. Ask the same question about different embedding levels, so you can tell real structural errors apart from shortcuts that rely on linear adjacency between a noun and a verb.
  • I3. Treat a zero-shot failure as possibly a problem of understanding the task rather than a lack of knowledge; demonstrations can reveal what a model really knows.
  • I4. Classify errors by which wrong noun the model chose (adjacent noun, another embedded noun, other) instead of only counting accuracy.
  • I5. Use ungrammatical variants with a missing verb (the "VP illusion" cases) as minimal-pair foils in grammaticality judgment.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (tree): Generate nested relative-clause sentences of levels 1–4 and check that the perceptron parser and gparse attach nsubj/obj and the relative-clause links correctly at every depth, as a structural stress test.
  • I2 (frames): When testing who-did-what-to-whom frame extraction, query roles at each nesting depth separately and report the results per depth, to show whether role assignment follows the tree or just word order.
  • I3 (learn): Before rejecting an LLM teacher tier, or trusting it as a silver-data witness, run it with a few worked examples, so task misreading is not confused with ignorance.
  • I4 (learn): Add a small error taxonomy of this kind to parser and reader evaluations, so self-induced repair rules can target a specific failure pattern such as adjacency attraction.
  • I5 (learn): Include missing-VP center-embedding sentences as negative controls; a sound pipeline should flag them rather than produce a confident tree and answer.

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

🎓Copenhagen Business SchoolCopenhagen · Denmark6 works · 53 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 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksEnglish · 2,263 works

Cite

Cite as Daniel Hardt. 2025. Sparks of Pure Competence in LLMs: the Case of Syntactic Center Embedding in English. Proceedings of the Society for Computation in Linguistics 2025.

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

@inproceedings{hardt2025sparks,
  title = {{Sparks of Pure Competence in LLMs: the Case of Syntactic Center Embedding in English}},
  author = {Daniel Hardt},
  booktitle = {Proceedings of the Society for Computation in Linguistics 2025},
  year = {2025},
  doi = {10.7275/scil.3149},
  url = {https://doi.org/10.7275/scil.3149}
}

References

24 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Steven P Abney and Mark Johnson. 1991. Memory requirements and local ambiguities of parsing strategies. Journal of Psycholinguistic Research, 20:233– 250. ↗
  2. Emmon Bach, Colin Brown, and William Marslen-Wilson. 1986. Crossed and nested dependencies in German and Dutch: A psycholinguistic study. Language and Cognitive Processes, 1(4):249–262. ↗
  3. Ted Chiang. 1998. Story of your life. Stories of your life and others, pages 117–78. find
  4. Ruixiang Cui, Seolhwa Lee, Daniel Hershcovich, and Anders Søgaard. 2023. What does the failure to reason with “respectively” in zero/few-shot settings tell us about language models? In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8786–8800, Toronto, Canada. Association for Computational Linguistics. find
  5. Vittoria Dentella, Fritz Günther, and Evelina Leivada. 2023. Systematic testing of three language models reveals low language accuracy, absence of response stability, and a yes-response bias. Proceedings of the National Academy of Sciences, 120(51):e2309583120. ↗
  6. Gabe Dupre. 2021. (what) can deep learning contribute to theoretical linguistics? Minds and Machines, 31(4):617–635. ↗
  7. Felix Engelmann and Shravan Vasishth. 2009. Processing grammatical and ungrammatical center embeddings in English and German: A computational model. In Proceedings of the Ninth International Conference on Cognitive Modeling, Manchester, UK, pages 240–45. find
  8. Jon Gauthier, Jennifer Hu, Ethan Wilcox, Peng Qian, and Roger Levy. 2020. SyntaxGym: An online platform for targeted evaluation of language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 70–76, Online. Association for Computational Linguistics.
  9. Edward Gibson. 1998. Linguistic complexity: Locality of syntactic dependencies. Cognition, 68(1):1–76. ↗
  10. Edward Gibson and James Thomas. 1999. Memory limitations and structural forgetting: The perception of complex ungrammatical sentences as grammatical. Language and Cognitive Processes, 14(3):225–248. ↗
  11. Daniel Hardt. 2023. Ellipsis-dependent reasoning: a new challenge for large language models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 39–47. Association for Computational Linguistics. ↗
  12. The Jennifer Hu and Michael C Frank. 2024. Auxiliary task demands mask the capabilities of smaller language models. arXiv preprint arXiv:2404.02418. ↗
  13. Noam Chomsky. 1965. Aspects of the Theory of Syntax. 11. MIT press. ↗
  14. Jennifer Hu and Roger Levy. 2023. Prompting is not a substitute for probability measurements in large language models. arXiv preprint arXiv:2305.13264. ↗
  15. Noam Chomsky. 1957. Hague: Mouton. Syntactic structures. find
  16. Noam Chomsky, George Armitage Miller, R Luce, R Bush, and E Galanter. 1963. Introduction to the formal analysis of natural languages. 1963, pages 269–321. find
  17. Morten H Christiansen. 1997. The (un) grammaticality of doubly center-embedded sentences: a connectionist perspective. In Poster presented at the 10th CUNY Sentence Processing Conference, Santa Monica, CA. find
  18. Jennifer Hu, Kyle Mahowald, Gary Lupyan, Anna Ivanova, and Roger Levy. 2024. Language models align with human judgments on key grammatical constructions. Proceedings of the National Academy of Sciences, 121(36):e2400917121. ↗
  19. Fred Karlsson. 2007. Constraints on multiple center-embedding of clauses. Journal of Linguistics, 43(2):365–392. ↗
  20. Thomas S Kuhn. 1997. The structure of scientific revolutions, volume 962. University of Chicago press Chicago. find
  21. Yair Lakretz, Dieuwke Hupkes, Alessandra Vergallito, Marco Marelli, Marco Baroni, and Stanislas Dehaene. 2021. Mechanisms for handling nested dependencies in neural-network language models and humans. Cognition, 213:104699. Special Issue in Honour of Jacques Mehler, Cognition’s founding editor. ↗
  22. Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg. 2016. Assessing the ability of LSTMs to learn syntax-sensitive dependencies. Transactions of the Association for Computational Linguistics, 4:521–535. ↗
  23. Kyle Mahowald, 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. ↗
  24. George A Miller and Stephen Isard. 1964. Free recall of self-embedded English sentences. Information and control, 7(3):292–303. period, also with default settings. Each of these tests were performed with 200 randomly selected examples. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/hardt-2025-sparks-pure-competence-llms-case/