Large language models have features, such as hidden layers, that implicitly build in some insights of linguistics.

Judge against interpretability and probing research available by 2025-09-24, such as structural-probe studies showing that syntax trees, parts of speech, agreement or hierarchical structure are encoded in transformer hidden layers. True: peer-reviewed studies show that language models' internal representations encode linguistic structures or categories recognised in linguistics. Mostly true: such encoding is shown, but it is partial or approximate in a way the claim did not depend on. Mixed: the findings are seriously contested, with comparable evidence that the apparent encodings are artefacts of the probing method. Mostly false: only weak or isolated evidence exists. False: the research shows that models do not represent linguistic structure.

Set before the research began.

1. E1 A Structural Probe for Finding Syntax in Word Representations Association for Computational Linguistics (NAACL 2019 proceedings) · June 2019

   “Using our probe, we show that such transformations exist for both ELMo and BERT but not in baselines, providing evidence that entire syntax trees are embedded implicitly in deep models' vector geometry.”

2. E2 Emergent linguistic structure in artificial neural networks trained by self-supervision PNAS (hosted on PubMed Central) · 3 June 2020

   “we show that a linear transformation of learned embeddings in these models captures parse tree distances to a surprising degree, allowing approximate reconstruction of the sentence tree structures normally assumed by linguists.”

3. E3 BERT Rediscovers the Classical NLP Pipeline Association for Computational Linguistics (ACL 2019 proceedings) · 4 July 2019

   “We find that the model represents the steps of the traditional NLP pipeline in an interpretable and localizable way, and that the regions responsible for each step appear in the expected sequence: POS tagging, parsing, NER, semantic roles, then coreference.”

4. E4 Does BERT Rediscover a Classical NLP Pipeline? International Committee on Computational Linguistics (COLING 2022 proceedings) · 4 October 2022

   “In re-examining Jawahar et al. (2019) and Tenney et al.’s (2019a) probes into the structure of BERT, we have found that the pipeline-like separation that they asserted lacks conclusive empirical support.”

Peer-reviewed studies from before 2025-09-24 show that language models' hidden representations encode structures recognised in linguistics. E1 finds evidence that whole syntax trees are embedded in the vector geometry of ELMo and BERT. E2 (PNAS) finds that a linear transformation of model embeddings allows approximate reconstruction of linguists' parse trees. E3 finds that BERT encodes POS tagging, parsing and semantic roles in interpretable, localisable regions. E4 finds that the claimed pipeline-like separation of those regions lacks conclusive empirical support, and E2 calls the reconstruction approximate, so the encoding is shown but is partial in a way the claim does not depend on.