Elamur Research⌕
☰
home / papers / 2026 / Do large language models and humans follow similar learning stages? Assessing GPT-2's order of Swedish grammar acquisition within the Processability Theory framework

Do large language models and humans follow similar learning stages? Assessing GPT-2's order of Swedish grammar acquisition within the Processability Theory framework

Paper · 2026 · Stella Lundqvist, Murathan Kurfali and Johan Sjons

Summary

The paper asks whether a language model learning a new language picks up grammar in the same stepwise order that Processability Theory (PT) predicts for human second-language learners. To simulate an L2 learner, the authors took an English-pretrained GPT-2 and fine-tuned it on Swedish text from OSCAR. The Swedish data was parsed with Stanza and sorted by PT developmental stage. It was then presented in three curricula: random order, increasing complexity and decreasing complexity. They also built SwePT, a new minimal-pair benchmark extracted from the UD Swedish Talbanken and LinES treebanks. SwePT covers nine grammatical phenomena spanning PT stages 2–5. At regular checkpoints the models took acceptability judgment tests (AJT). These were scored with adapted versions of the emergence criterion and implicational scaling. The acquisition order was largely the same across curricula, but it did not follow PT's predicted hierarchy. Performance seemed driven mostly by unigram and n-gram frequency heuristics. The study brings SLA theory into LLM evaluation, though the authors caution that PT and AJT-based testing are not fully compatible.

Key points

  • SwePT benchmark: nine subsets mapped to PT stages.
    • Stage 2: SVO, PLU, TNS
    • Stage 3: ATT
    • Stage 4: PR_a, PR_b, INV
    • Stage 5: NEG, INQ
    • Data and code are released at github.com/stellson/SwePT.
  • How the minimal pairs were built: sentences were extracted by rule-based dependency-tree search over UD CoNLL-U files. Syntactic items were perturbed by reordering constituents and morphological items by reverting forms to the lemma. The PR_b pairs were altered by hand.
  • Fine-tuning data: about 9% of Swedish OSCAR (roughly 680k examples, ~1B tokens) was parsed with Stanza and labeled by structure. A 500k-sentence sample was used so that one epoch fit within 72 hours.
  • Main result: rank correlations show the acquisition order is robust to curriculum ordering, but it violates PT's implicational sequence.
  • Limitations and future work: binary accuracy fits poorly with PT's categorical emergence criterion, and minimal pairs do not control for length or frequency. The authors suggest analyzing model production instead and adding a Swedish model such as GPT-SW3.

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

🎓Uppsala UniversityUppsala · Sweden64 works · 13,164 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 worksPsychology · 1,552 worksDevelopmental and Educational Psychology · 499 works

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

Cite

Cite as Stella Lundqvist, Murathan Kurfali and Johan Sjons. 2026. Do large language models and humans follow similar learning stages? Assessing GPT-2's order of Swedish grammar acquisition within the Processability Theory framework. Proceedings of the 1st Workshop on Computational Developmental Linguistics (CDL).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.cdl-1.10

@inproceedings{lundqvist2026do,
  title = {{Do large language models and humans follow similar learning stages? Assessing GPT-2's order of Swedish grammar acquisition within the Processability Theory framework}},
  author = {Stella Lundqvist, Murathan Kurfali and Johan Sjons},
  booktitle = {Proceedings of the 1st Workshop on Computational Developmental Linguistics (CDL)},
  year = {2026},
  doi = {10.18653/v1/2026.cdl-1.10},
  url = {https://doi.org/10.18653/v1/2026.cdl-1.10}
}

References

8 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Gerry TM Altmann and Yuki Kamide. 1999. Incremental interpretation at verbs: Restricting the domain of subsequent reference. Cognition, 73(3):247–264. ↗
  2. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41–48.
  3. Terra Blevins, Hila Gonen, and Luke Zettlemoyer. 2022. Analyzing the mono-and cross-lingual pretraining dynamics of multilingual language models. arXiv preprint arXiv:2205.11758. ↗
  4. Aafke Buyl and Alex Housen. 2015. Developmental stages in receptive grammar acquisition: A processability theory account. Second Language Research, 31(4):523–550. ↗
  5. Cheng-Han Chiang, Sung-Feng Huang, and Hung-yi Lee. 2020. Pretrained language model embryology: The birth of albert. arXiv preprint arXiv:2010.02480. ↗
  6. Leshem Choshen, Guy Hacohen, Daphna Weinshall, and Omri Abend. 2021. The grammar-learning trajectories of neural language models. arXiv preprint arXiv:2109.06096. ↗
  7. Marie-Catherine De Marneffe, Christopher D Manning, Joakim Nivre, and Daniel Zeman. 2021. Universal dependencies. Computational linguistics, 47(2):255– 308.
  8. Bronwen Patricia Dyson and Gisela Håkansson. 2017. Understanding second language processing: A focus on Processability Theory, volume 4. John Benjamins Publishing Company. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lundqvist-2026-do-large-language-models-humans/