Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

How Syntax Specialization Emerges in Language Models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Syntactic specialization in language models is not present at initialization but emerges gradually, concentrates in specific layers, and passes through a critical period around 16 million tokens of training.

desk verdict A new metric and broad developmental survey, but the headline 'critical period' is contradicted by the paper's own late-phase divergence statistic. read the letter →

arxiv 2505.19548 v1 pith:CQJ67EBL submitted 2025-05-26 cs.CL cs.AI

Xufeng Duan , Zhaoqian Yao , Yunhao Zhang , Shaonan Wang , Zhenguang G. Cai This is my paper
classification cs.CLcs.AI
keywords syntacticspecializationSensitivityIndexlanguagemodelinterpretabilitytrainingdynamicscriticalperiodminimalpairsBLiMPscaleandinitialization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tracks how transformer language models come to represent syntax internally, using a new classifier-free measure, the Syntactic Sensitivity Index (SSI), computed from activation differences between grammatical and ungrammatical minimal pairs. Across training checkpoints of retrained GPT-2 models and the Pythia suite, the authors find that syntactic sensitivity starts near zero, rises gradually, concentrates in specific layers, and passes through a critical period in the first roughly 16 million tokens, after which models with different random seeds converge on the same layer-level profile. They also report that scaling the model increases and sharpens syntactic abstraction, that training-data size shifts it per phenomenon, and that the specific neurons carrying syntax differ across seeds even when the layer profile does not. This turns 'where syntax lives in an LLM' from a static localization result into a developmental story with a measured timescale, testable on the released checkpoints.

What carries the argument

The load-bearing object is the Syntactic Sensitivity Index (SSI), defined per syntactic phenomenon $p$ and layer $l$ as $\mathrm{SSI}^{(l)}_p = \mathrm{Intra}^{(l)}_p - \mathrm{Inter}^{(l)}_p$. For each minimal pair from BLiMP, the model produces a difference vector $\Delta h = h_g - h_u$ from mean-pooled, L2-normalized layer activations of the grammatical and ungrammatical sentences; Intra is the average cosine similarity among the $\Delta h$ vectors within a phenomenon, and Inter is the average cosine similarity to the vectors of all other phenomena. High SSI means the model's internal grammatical/ungrammatical contrast is consistent inside a phenomenon and distinct from other phenomena, which the paper treats as the operational signature of syntactic specialization. The same statistics are pushed to the neuron level by correlating each feedforward-layer dimension's scalar response across all pairs and z-scoring it against a background of other phenomena, flagging neurons in the top 25% correlation with a z-score above 2 as syntax-sensitive.

What would settle it

Compute SSI on control pairs in which each ungrammatical BLiMP sentence is replaced by a second grammatical sentence matched for length and word frequency, so the pairs carry no syntactic contrast; a syntax-tracking index should sit near zero at every checkpoint, and reproducing the reported SSI values would show the metric measures sentence-pair dissimilarity instead. For the critical-period claim specifically, train a third seed and swap its training corpus immediately after 16 million tokens: if layer-level SSI stays pinned to the first corpus, the window is a genuine lock-in, and if it tracks the new corpus, the convergence finding is an artifact of corpus composition.

Watch

Extended reading notes

Core claim

The paper's central claim is that internal syntactic specialization in transformer language models is not present at initialization but develops along a staged trajectory: it emerges gradually as the model sees data, it concentrates in specific layers, and it passes through a critical period early in training — models initialized with different random seeds converge on similar layer-level syntactic representations after approximately 16 million tokens. The claim is supported by three converging lines of evidence: SSI tracks grammaticality-judgment accuracy across checkpoints in linear mixed-effects models ($\beta = 0.15$, $t = 9.50$ for GPT-2), whereas supervised SVM and regression probes show no significant change from untrained to trained models; ablating high-SSI neurons degrades perplexity far more than random ablation (a mean increase of 631 points for GPT-2, 16,414 for Pythia); and layer-level SSI profiles correlate at 0.98 across seeds of the same architecture versus 0.66 across architectures. A companion finding is the convergent-layer/divergent-neuron split: where syntax specializes is reproducible across seeds, but which neurons implement it is not, with only about 1.5% overlap in top-SSI neuron identity.

Load-bearing premise

The central assumption is that the activation difference between a grammatical sentence and its ungrammatical twin is a clean signature of syntax rather than of surface confounds such as word frequency, sentence length, or lexical overlap, and the paper reports no control analysis that would rule those out.

Editorial extensions

If this is right

  • SSI provides a classifier-free monitor of syntactic competence, so the emergence of syntax can be tracked checkpoint by checkpoint during pretraining without behavioral probes or supervised probes.
  • The roughly 16-million-token convergence window implies that the earliest portion of pretraining data, not later training, determines the layer-level placement of syntactic knowledge, which could inform curriculum or data-ordering interventions.
  • Because layer-level specialization is reproducible across seeds but neuron-level circuits are not, layer-level SSI profiles are the stable unit for comparing models, while neuron maps must be treated as instance-specific.
  • Model scale monotonically increases and layer-localizes syntactic abstraction across the Pythia series (70M to 1.4B parameters), so scaling laws apply to internal syntactic structure and not only to loss.
  • Different syntactic phenomena follow different acquisition schedules — ellipsis differentiates earliest and in lower layers, determiner–noun agreement latest and in upper layers — so syntactic development is not a single event but a family of trajectories.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct intervention test of the critical-period claim follows from the paper's own setup: train additional seeds and swap the training corpus just after 16 million tokens; if layer-level SSI follows the new corpus, the window is not a lock-in, and if it stays pinned, the critical-period reading is supported. The authors do not run this experiment.
  • The paper reports no confound control for SSI, so a natural next check is grammatical-versus-grammatical control pairs; if SSI on pairs without a syntactic contrast approaches the reported values, the metric would need recalibration before the developmental conclusions hold.
  • The convergent-layer/divergent-neuron split suggests that averaging or ensembling models across seeds would wash out neuron-level syntax maps, which matters for any downstream work that proposes to edit or prune syntax circuitry.
  • Extending SSI from mean-pooled sentence embeddings to token-level activation differences would test whether the critical period is visible at individual syntactic positions, since mean pooling could smooth away positional specialization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces the Syntactic Sensitivity Index (SSI), an unsupervised metric that measures how consistently a model's internal activations separate grammatical from ungrammatical BLiMP minimal pairs. Using SSI, the authors track syntactic specialization across GPT-2 and Pythia training checkpoints and report that (i) SSI correlates with grammaticality-judgment accuracy, (ii) ablating high-SSI neurons increases perplexity, (iii) syntactic sensitivity emerges gradually and concentrates in specific layers, (iv) models with different seeds converge on layer-level specialization after roughly 16 million tokens, consistent with a 'critical period', and (v) model scale and training data modulate specialization. The paper promises release of code, models, and checkpoints.

Significance. If the central claims held, the SSI would be a useful unsupervised tool for studying the development of syntactic representations, and the longitudinal checkpoint analysis would be a valuable contribution to the interpretability literature. The authors also report a large set of training checkpoints and plan to release artifacts, which is a genuine strength. However, the headline 'critical period' claim is directly contradicted by the paper's own reported statistics, and the metric validation is entirely in-sample. As submitted, the evidence does not support the abstract's central conclusion, and the developmental trajectory results rest on an unvalidated and possibly confounded measure.

major comments (4)
  1. [§4.4 and Appendix A.4] The statistical result reported in the section that is supposed to establish the critical-period claim has the opposite sign from what the claim requires. The text reports 'greater SSI divergence in the late vs. early phase (0.19 vs. −0.02, β = 0.229, t = 8.77, p < .001)' and Appendix A.4 repeats that the late phase shows a 'significant increase in ∆SSI'. This means seed divergence is larger late in training, not convergent after ~16M tokens. The following paragraph nonetheless asserts that 'these differences diminished and eventually converged'. This is an internal contradiction, and Figure 6 is not accompanied by any quantitative convergence test. The abstract and Figure 1(b) therefore make a claim that the only quantitative evidence in the paper contradicts.
  2. [§3.1, §4.1.1] The validation of SSI is in-sample: the same BLiMP minimal pairs are used to compute SSI, grammaticality accuracy, and the perplexity-based ablation outcome. The correlation between ΔSSI and ΔAccuracy therefore does not independently confirm that SSI measures syntactic competence; it only shows that the metric is related to a behavior measured on the same items. An out-of-sample or cross-benchmark validation, or control analyses using non-syntactic sentence pairs, would be needed to support the claim that SSI captures syntax rather than item-specific surface properties.
  3. [§3.1] The SSI definition uses the mean-pooled activation difference Δh between a grammatical sentence and its ungrammatical counterpart. BLiMP minimal pairs differ by a single word, so Δh can be driven by lexical identity, word frequency, sentence length, or other non-syntactic differences between the two sentences. The paper reports no control analyses for such confounds, even though every downstream finding—the developmental trajectory, the seed-convergence result, and the scale/data effects—depends on interpreting SSI as a measure of syntactic structure. Without such controls, the central construct validity is not established.
  4. [§4.4, Figure 6] Even apart from the sign contradiction, the 'approximately 16 million tokens' boundary is not derived from any formal test. It appears to be a visual reading of Figure 6, and the mixed-effects model only compares two arbitrarily defined phases (≤16M vs. >16M tokens). With only two GPT-2 seeds and no model of convergence time, the data are insufficient to support a 'critical period' claim of the kind made in the abstract and Section 5.
minor comments (6)
  1. [§3.2 vs. §4.1.2] Section 3.2 states that syntax-sensitive neurons are those in the 'top 25% correlation' with a z-score threshold of 2, but Section 4.1.2 reports that selected neurons correspond to 'the top 5% of 9,216 total neurons'. These thresholds are inconsistent and should be reconciled.
  2. [Figure 1(b)] The comparison between human neural plasticity and model SSI divergence uses different quantities and different y-axes. The visual analogy may overstate the parallel; the figure caption should clarify that the human curve is not a direct measurement from this study and that the model curve is not on the same scale.
  3. [§3.6] The checkpoint description '20, 21, 22, . . . ,211 million tokens' is ambiguous; it should be written as powers of two (2^0, 2^1, ..., 2^11 million tokens) to match the values 0, 2, 4, ..., 2048 used in Appendix A.1.
  4. [§4.2] The sentence ending '(? )' appears to be an editorial artifact and should be removed or completed.
  5. [Appendix A.4] The normalization of ΔSSI is described as 'normalized by the mean SSI of the two seeds at each data point', but it is not specified whether the absolute difference is divided by the mean or by some other quantity; the exact formula should be given.
  6. [Appendix A.1] The likelihood-ratio tests use α = 0.2, which is an unusual significance level; the authors should either justify this choice or use a conventional level such as 0.05.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular reduction found: SSI is an operational measure and the developmental/validation analyses are empirical, not definitional. Main correctness caveat: the only quantitative test for the critical-period claim (Appendix A.4) has the opposite sign from the claim.

full rationale

The paper's derivation chain is empirical rather than circular. SSI (Sec. 3.1) is defined as Intra-group minus Inter-group cosine similarity of activation differences on BLiMP minimal pairs; the developmental claim (Sec. 4.2) is that this quantity increases with training tokens (beta = 0.006, t = 15.17, p < .001). Accuracy is computed independently from mean log-probabilities (Sec. 3.4), so SSI does not define accuracy by construction. The abstract's 'SSI significantly predicts the accuracy' language (Sec. 4.1.1) is a concurrent mixed-effects correlation between scaled DeltaSSI and DeltaAccuracy at the same checkpoints; both variables are measured as absolute deviations from the final checkpoint, so shared time-trends inflate the association. This is in-sample validation, not a fitted parameter renamed as a prediction, and it is not an equality by construction. The ablation analysis (Sec. 4.1.2) also uses the same grammatical sentences for neuron selection and PPL measurement, but the random-ablation control makes the comparison nontrivial. Self-citations (Duan et al. 2025; Qiu et al. 2025) appear in related work and are not load-bearing for SSI or the developmental findings. A separate, non-circularity problem is that the critical-period claim (Sec. 4.4) is contradicted by the paper's own Appendix A.4: the mixed-effects model reports a significant increase in normalized SSI divergence in the late vs. early phase (0.19 vs. -0.02, beta = 0.229, t = 8.77, p < .001), i.e., greater divergence later, whereas the text claims convergence after ~16M tokens. This should be corrected, but it is an internal inconsistency, not a circular step.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the validity of BLiMP as a syntactic benchmark, the assumption that mean-pooled activations capture syntactic structure, and the hand-selected thresholds in the SSI definition and the critical period boundary. The unusual training corpus composition (multilingual parallel data) is an additional unstated assumption.

free parameters (4)
  • top 25% correlation threshold = 0.25
    Used in Section 3.2 to select syntax-sensitive neurons; hand-chosen, not fitted to data.
  • z-score threshold = 2
    Used in Section 3.2 neuron selection; conventional but arbitrary.
  • critical period boundary = 16M tokens
    Used in Section 4.4 to group early and late training phases; hand-selected and load-bearing for the critical period claim.
  • alpha for likelihood ratio tests = 0.2
    Used in Appendix A.1 and A.2; unusually high significance threshold, inflating support for random slopes.
assumptions (4)
  • domain assumption BLiMP minimal pairs accurately reflect syntactic grammaticality
    The paper uses BLiMP as ground truth for grammatical vs ungrammatical in all SSI, accuracy, and ablation analyses (Section 3.3).
  • domain assumption Mean-pooled sentence embeddings preserve syntactic information
    SSI is computed on mean-pooled and L2-normalized activations; the paper's own limitations section notes this may miss token-level nuance.
  • standard math The L2 normalization and cosine similarity are appropriate for comparing activation differences
    Implicit in the SSI definition (Section 3.1).
  • domain assumption The training corpora are representative of English text for syntax acquisition
    The small-scale corpus includes multilingual parallel corpora (UNCorpus, translation2019zh, WikiMatrix), which conflicts with the framing of English LMs (Appendix A.6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Syntax Specialization Emerges in Language Models." pith.science (2026). https://pith.science/paper/CQJ67EBL

@misc{pith2026250519548,
  author       = {Pith},
  title        = {Pith review of: How Syntax Specialization Emerges in Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CQJ67EBL}},
  note         = {Machine review of arXiv:2505.19548}
}
read the original abstract

Large language models (LLMs) have been found to develop surprising internal specializations: Individual neurons, attention heads, and circuits become selectively sensitive to syntactic structure, reflecting patterns observed in the human brain. While this specialization is well-documented, how it emerges during training and what influences its development remains largely unknown. In this work, we tap into the black box of specialization by tracking its formation over time. By quantifying internal syntactic consistency across minimal pairs from various syntactic phenomena, we identify a clear developmental trajectory: Syntactic sensitivity emerges gradually, concentrates in specific layers, and exhibits a 'critical period' of rapid internal specialization. This process is consistent across architectures and initialization parameters (e.g., random seeds), and is influenced by model scale and training data. We therefore reveal not only where syntax arises in LLMs but also how some models internalize it during training. To support future research, we will release the code, models, and training checkpoints upon acceptance.

Figures

Figures reproduced from arXiv: 2505.19548 by the authors.

Figure 1
Figure 1. (a) Left. Differences in SSI and grammatical judgment task accuracy between various checkpoints during [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Specialized language related regions from ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The neuron ablation result with 3 SD of origi [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: Initialization Analysis. Each line represents [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Model scale analysis. Each line represents a [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Training data analysis. SSI layer-wise trajec [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Developmental trajectories of SSI across syntactic phenomena in GPT-2 (seed 1 and seed 7). [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: SSI trajectories across seeds and architectures for ten syntactic phenomena. Each line represents a model [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Each line represents a model from the Pythia series (70M–1.4B parameters), with SSI plotted across [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Layer-wise SSI trajectories are shown for multiple syntactic phenomena in GPT-2 models trained on [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Linear representations of grammaticality in neural language models

    cs.CL 2026-07 conditional novelty 5.0 of 10

    Grammaticality is linearly decodable from language model sentence representations and generalizes across phenomena and languages in larger models.

  2. The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models

    cs.CL 2026-01 conditional novelty 4.0 of 10

    A systematic review of 337 articles shows Transformers handle formal syntax well but perform worse and more variably at the syntax-semantics interface, with the field over-reliant on English and BERT.

Reference graph

Works this paper leans on

41 extracted references · 20 canonical work pages · cited by 2 Pith papers

  1. [1]

    Badr AlKhamissi, Greta Tuckute, Yingtian Tang, Taha Binhuraib, Antoine Bosselut, and Martin Schrimpf. 2025. From language to cognition: How llms outgrow the human language network. arXiv preprint arXiv:2503.01830

    arXiv 2025

  2. [2]

    Emmanuel Ameisen, Jack Lindsey, Adam Pearce, Wes Gurnee, Nicholas L. Turner, Brian Chen, Craig Citro, David Abrahams, Shan Carter, Basil Hosmer, Jonathan Marcus, Michael Sklar, Adly Templeton, Trenton Bricken, Callum McDougall, Hoagy Cunningham, Thomas Henighan, Adam Jermyn, Andy Jones, and 8 others. 2025. https://transformer-circuits.pub/2025/attribution...

    2025

  3. [3]

    Douglas Bates, Martin M \"a chler, Ben Bolker, and Steve Walker. 2015. Fitting linear mixed-effects models using lme4. Journal of statistical software, 67:1--48

    work page 2015

  4. [4]

    Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar van der Wal. 2023. https://arxiv.org/abs/2304.01373 Pythia: A suite for analyzing large language models across training and scaling . Preprint...

    arXiv 2023

  5. [5]

    Jannik Brinkmann, Chris Wendler, Christian Bartelt, and Aaron Mueller. 2025. Large language models share representations of latent grammatical concepts across typologically diverse languages. arXiv preprint arXiv:2501.06346

    arXiv 2025

  6. [6]

    Angelica Chen, Ravid Shwartz-Ziv, Kyunghyun Cho, Matthew L Leavitt, and Naomi Saphra. 2024. https://openreview.net/forum?id=MO5PiKHELW Sudden drops in the loss: Syntax acquisition, phase transitions, and simplicity bias in MLM s . In The Twelfth International Conference on Learning Representations

    work page 2024

  7. [7]

    Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. https://doi.org/10.18653/v1/W19-4828 What does BERT look at? an analysis of BERT `s attention . In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276--286, Florence, Italy. Association for Computational Linguistics

    doi:10.18653/v1/w19-4828 2019

  8. [8]

    Susan Curtiss. 1977. Genie: A Psycholinguistic Study of a Modern-Day Wild Child. Academic Press, New York

    work page 1977

Show all 41 references
  1. [9]

    Xufeng Duan, Xinyu Zhou, Bei Xiao, and Zhenguang Cai. 2025. Unveiling language competence neurons: A psycholinguistic approach to model interpretability. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10148--10157

    2025

  2. [10]

    Evelina Fedorenko and Nancy Kanwisher. 2009. Neuroimaging of language: why hasn't a clearer picture emerged? Language and Linguistics Compass, 3(4):839--865

    2009

  3. [11]

    Evelina Fedorenko, Steven T Piantadosi, and Edward AF Gibson. 2024. Language is primarily a tool for communication rather than thought. Nature, 630(8017):575--586

    2024

  4. [12]

    Angela D Friederici. 2018. The neural basis for human syntax: Broca's area and beyond. Current opinion in behavioral sciences, 21:88--92

    2018

  5. [13]

    Hall, Dongmei Li, and Tim D

    Wyatte C. Hall, Dongmei Li, and Tim D. V. Dye. 2018. https://doi.org/10.2105/AJPH.2018.304498 Influence of hearing loss on child behavioral and home experiences . American Journal of Public Health, 108(8):1079--1081. Epub 2018 Jun 21

    2018

  6. [14]

    Hertz-Pannier, C

    L. Hertz-Pannier, C. Chiron, I. Jambaque, V. Renaux-Kieffer, P. F. Van de Moortele, O. Delalande, M. Fohlen, F. Brunelle, and D. Le Bihan. 2002. Late plasticity for language in a child's non-dominant hemisphere: a pre- and post-surgery fmri study. Brain, 125(2):361--372

    2002

  7. [15]

    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

    2024

  8. [16]

    Johnson and Elissa L

    Jacqueline S. Johnson and Elissa L. Newport. 1989. Critical period effects in second language learning: The influence of maturational state on the acquisition of english as a second language. Cognitive Psychology, 21(1):60--99

    1989

  9. [17]

    MS Keshavan, J Giedd, JY Lau, DA Lewis, and T Paus. 2014. https://doi.org/10.1016/S2215-0366(14)00081-9 Changes in the adolescent brain and the pathophysiology of psychotic disorders . Lancet Psychiatry, 1(7):549--558

    2014 doi

  10. [18]

    Hassane Kissane, Achim Schilling, and Patrick Krauss. 2025. Probing internal representations of multi-word verbs in large language models. arXiv preprint arXiv:2502.04789

    2025 arXiv

  11. [19]

    Cheslie C Klein, Philipp Berger, Tom \'a s Goucha, Angela D Friederici, and Charlotte Grosse Wiesmann. 2023. Children’s syntax is supported by the maturation of ba44 at 4 years, but of the posterior sts at 3 years of age. Cerebral Cortex, 33(9):5426--5435

    2023

  12. [20]

    Alexandra Kuznetsova, Per B Brockhoff, and Rune HB Christensen. 2017. lmertest package: tests in linear mixed effects models. Journal of statistical software, 82:1--26

    2017

  13. [21]

    Yair Lakretz, German Kruszewski, Theo Desbordes, Dieuwke Hupkes, Stanislas Dehaene, and Marco Baroni. 2019. https://doi.org/10.18653/v1/N19-1002 The emergence of number and syntax units in LSTM language models . In Proceedings of the 2019 Conference of the North American Chapt...

    2019 doi

  14. [22]

    Lenneberg

    Eric H. Lenneberg. 1967. Biological Foundations of Language. Wiley, New York

    1967

  15. [23]

    Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg. 2016. https://doi.org/10.1162/tacl_a_00115 Assessing the ability of LSTM s to learn syntax-sensitive dependencies . Transactions of the Association for Computational Linguistics, 4:521--535

    2016 doi

  16. [24]

    Yiheng Liu, Xiaohui Gao, Haiyang Sun, Bao Ge, Tianming Liu, Junwei Han, and Xintao Hu. 2025. Brain-inspired exploration of functional networks and key neurons in large language models. arXiv preprint arXiv:2502.20408

    2025

  17. [25]

    Yikang Liu, Yeting Shen, Hongao Zhu, Lilong Xu, Zhiheng Qian, Siyuan Song, Kejia Zhang, Jialong Tang, Pei Zhang, Baosong Yang, and 1 others. 2024. Zhoblimp: a systematic assessment of language models with linguistic minimal pairs in chinese. arXiv preprint arXiv:2411.06096

    2024

  18. [26]

    Rebecca Marvin and Tal Linzen. 2018. Targeted syntactic evaluation of language models. arXiv preprint arXiv:1808.09031

    2018 arXiv

  19. [27]

    Aaron Mueller, Yu Xia, and Tal Linzen. 2022. Causal analysis of syntactic agreement neurons in multilingual language models. arXiv preprint arXiv:2210.14328

    2022 arXiv

  20. [28]

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, and 1 others. 2022. In-context learning and induction heads. arXiv preprint arXiv:2209.11895

    2022 arXiv

  21. [29]

    S Pénicaud, D Klein, RJ Zatorre, JK Chen, P Witcher, K Hyde, and RI Mayberry. 2013. https://doi.org/10.1016/j.neuroimage.2012.09.076 Structural brain changes linked to delayed first language acquisition in congenitally deaf individuals . Neuroimage, 66:42--49

    2013 doi

  22. [30]

    Zhuang Qiu, Xufeng Duan, and Zhenguang G Cai. 2025. Grammaticality representation in chatgpt as compared to linguists and laypeople. Humanities and Social Sciences Communications, 12(1):1--15

    2025

  23. [31]

    Alec Radford, Jeffrey Wu, Rewon Child, and 1 others. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1. Available at https://openai.com/blog/better-language-models/

    2019

  24. [32]

    Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. 2017. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems, 30

    2017

  25. [33]

    Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Xin Zhao, Furu Wei, and Ji-Rong Wen. 2024. Language-specific neurons: The key to multilingual capabilities in large language models. arXiv preprint arXiv:2402.16438

    2024 arXiv

  26. [34]

    Oskar van der Wal, Pietro Lesci, Max M \"u ller-Eberstein, Naomi Saphra, Hailey Schoelkopf, Willem Zuidema, and Stella Biderman. 2025. Polypythias: Stability and outliers across fifty language model pre-training runs. In The Thirteenth International Conference on Learning Repr...

    2025

  27. [35]

    Vicari, A

    S. Vicari, A. Albertoni, A. M. Chilosi, P. Cipriani, G. Cioni, and E. Bates. 2000. Plasticity and reorganization during language development in children with early brain injury. Cortex, 36(1):31--46

    2000

  28. [36]

    Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. 2019. https://doi.org/10.18653/v1/P19-1580 Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned . In Proceedings of the 57th Annual Meeting of the Associatio...

    2019 doi

  29. [37]

    Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. 2022. Interpretability in the wild: a circuit for indirect object identification in gpt-2 small. arXiv preprint arXiv:2211.00593

    2022 arXiv

  30. [38]

    Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R. Bowman. 2020. https://doi.org/10.1162/tacl_a_00321 BL i MP : The benchmark of linguistic minimal pairs for E nglish . Transactions of the Association for Computational Linguistic...

    2020 doi

  31. [39]

    Hongchuan Zeng, Senyu Han, Lu Chen, and Kai Yu. 2025. https://aclanthology.org/2025.coling-main.707/ Converging to a lingua franca: Evolution of linguistic regions and semantics alignment in multilingual large language models . In Proceedings of the 31st International Conferen...

    2025

  32. [40]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.