pith. machine review for the scientific record. sign in

arxiv: 2512.03676 · v2 · submitted 2025-12-03 · 💻 cs.CL

Recognition: 2 theorem links

· Lean Theorem

Different types of syntactic agreement recruit the same units within large language models

Daria Kryvosheieva , Andrea de Varda , Evelina Fedorenko , Greta Tuckute
Authors on Pith no claims yet

Pith reviewed 2026-05-17 02:38 UTC · model grok-4.3

classification 💻 cs.CL
keywords syntactic agreementlarge language modelsfunctional localizationrepresentational overlapsubject-verb agreementcross-lingual syntaxgrammatical processinganaphor agreement
0
0 comments X

The pith

Different types of syntactic agreement recruit overlapping units inside large language models

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

The paper applies a functional localization method to pinpoint which units in seven open-weight LLMs respond most strongly to 67 syntactic phenomena. It reports that subject-verb, anaphor, and determiner-noun agreement consistently activate many of the same units, and that these units causally support the models' ability to distinguish grammatical from ungrammatical sentences. The same overlap appears in English, Russian, and Chinese, while a broader scan of 57 languages shows that structurally similar languages share more units for subject-verb agreement. A reader would care because the result suggests agreement forms a coherent functional category rather than a collection of unrelated rules inside the models.

Core claim

Using functional localization, the authors isolate units most responsive to each syntactic phenomenon and confirm that these units are reliably recruited across sentences and causally improve syntactic performance. Different agreement types recruit overlapping sets of units. This pattern is observed in English, Russian, and Chinese. In a cross-lingual analysis spanning 57 languages, structurally more similar languages share a larger proportion of units for subject-verb agreement.

What carries the argument

Functional localization procedure that selects LLM units showing the strongest response to a given syntactic phenomenon and tests whether intervening on those units affects grammatical judgments.

If this is right

  • Agreement processing in LLMs draws on shared representational resources instead of fully separate mechanisms for each subtype.
  • The same units that support one agreement relation can be expected to influence performance on other agreement relations.
  • Structurally similar languages will tend to align their subject-verb agreement units more closely than dissimilar languages do.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Model architectures could be designed with explicit agreement modules that leverage this shared representation.
  • Cross-lingual transfer of syntactic capabilities may be strongest between languages that already share agreement units.
  • Similar localization methods could be applied to other syntactic dependencies to test whether they also form functional categories.

Load-bearing premise

The units identified by the localization method are specifically involved in syntactic agreement rather than responding to correlated sentence features such as length, lexical items, or overall predictability.

What would settle it

If targeted interventions on the overlapping units impair only one agreement type while leaving the others intact, or if new models show no consistent overlap across agreement types, the claim that agreement forms a meaningful category would be undermined.

Figures

Figures reproduced from arXiv: 2512.03676 by Andrea de Varda, Daria Kryvosheieva, Evelina Fedorenko, Greta Tuckute.

Figure 1
Figure 1. Figure 1: Consistency of LLM units engaged across sentence instances for each syntactic phenomenon. Bars show overlap between unit sets identified in two independent halves of the data (2-fold cross-validation) for each of the 67 BLiMP phenomena, averaged across seven models. Bars are sorted from highest to lowest overlap (with percent overlap shown on the right); bar colors denote the category groupings in BLiMP an… view at source ↗
Figure 2
Figure 2. Figure 2: Performance impact of syntax-responsive unit ablation. Each colored bar shows the average dif￾ference in accuracy between the top-unit-ablated (1%) and unablated models for each of the 67 BLiMP phe￾nomena, averaged across seven models. Bars are sorted from highest to lowest top-unit ablation performance drop. Gray bars show accuracy differences from ablat￾ing a random 1% of units (averaged across seven mod… view at source ↗
Figure 4
Figure 4. Figure 4: Within-category and cross-category over￾laps in BLiMP. Each blue bar shows the average (across all pairs of phenomena within a category) intersection of localized unit sets (as a percentage out of the 1% target set; averaged across models). Each correspond￾ing orange bar shows the average intersection across all pairs of phenomena belonging to distinct categories. The categories are ordered by the differen… view at source ↗
Figure 3
Figure 3. Figure 3: Unit overlaps for all 2,211 pairs of BLiMP phenomena, averaged across seven models. Blue bars indicate unit overlaps for pairs of phenomena within the same syntactic category; orange bars indicate overlaps for pairs of phenomena across categories. The inset shows the pairs with the highest overlaps. The results are shown in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overlap of localized units within and across agreement categories. Bars show average overlap (as a percentage of top-1% units; averaged across models) between pairs of phenomena: (i) within each agreement category (blue; same data as in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Within-category and cross-category overlaps in Russian (RuBLiMP) and Chinese (SLING). The bars show average (across phenomena) within-category and cross-category overlaps for Russian (left) and Chinese (right), both using the Gemma model. The categories are ordered by the difference between the within-category and cross-category average overlap. ment phenomena = 9.99%; green bars in [PITH_FULL_IMAGE:figur… view at source ↗
Figure 7
Figure 7. Figure 7: Cross-lingual overlap of localized units for the agreement categories. For RuBLiMP (top) and SLING (bottom), bars show the average overlaps (as a percentage of the 1% top units; both use the Gemma model): (i) within each agreement category (blue); (ii) across different agreement categories within the same language (orange); (iii) between agreement and non￾agreement phenomena within the same language (gray)… view at source ↗
Figure 8
Figure 8. Figure 8: Relationship between language relatedness (syntactic similarity) and unit overlap, using the SV- # split from MultiBLiMP across 57 languages (1,586 pairs). Each dot represents a language pair; the x￾coordinate is the syntactic similarity between the two languages in the pair, and the y-coordinate is the unit overlap. The dotted line shows the line of best fit. Gemma model, we first investigated the range o… view at source ↗
Figure 9
Figure 9. Figure 9: Cross-validation analysis (2-fold) targeting [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ablation analysis targeting 0.5% of units. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Cross-phenomenon overlap analysis target [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Ablation analysis targeting 5% of units. [PITH_FULL_IMAGE:figures/full_fig_p015_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Cross-phenomenon overlap analysis target [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
Figure 17
Figure 17. Figure 17: Five-fold cross-validation analysis. E Generalization to other linguistic materials Do syntax-responsive LLM units identified from BLiMP generalize to other datasets targeting the same syntactic phenomena, or are they idiosyn￾cratic to BLiMP? To test this, we applied the same localization procedure to three additional minimal￾pair benchmarks: SyntaxGym (Gauthier et al., 2020), the naturalistic subset from… view at source ↗
Figure 16
Figure 16. Figure 16: Breakdown of syntax-responsive units (searched over the attention and MLP modules) by layer. Each row denotes a model; each bin in model m’s row denotes a layer in m. The x-axis indicates layer depth relative to the whole model (e.g., the 36th layer of a 48-layer model corresponds to x = 0.75). The color darkness indicates the percentage of unique units in the given layer out of the total unique units res… view at source ↗
Figure 18
Figure 18. Figure 18: Generalization of units localized on BLiMP to other materials. We compare unit overlap across benchmarks, within benchmarks, and the random baseline. Each orange bar shows the average cross￾benchmark overlap of the localized unit sets across all possible pairs of phenomena and all models; bench￾marks without phenomenon sub-divisions (Gulordava and Linzen) are treated as having one phenomenon. Blue bars de… view at source ↗
Figure 19
Figure 19. Figure 19: Results of the mean ablation experiment. [PITH_FULL_IMAGE:figures/full_fig_p017_19.png] view at source ↗
Figure 21
Figure 21. Figure 21: Scatterplot showing the relationship be￾tween cross-validation consistency and ablation ef￾fect of the localized unit sets. Each dot represents one of the 67 BLiMP phenomena; its x-coordinate is the average 2-fold cross-validation consistency of the respective top-1% unit set across the seven models con￾sidered, and its y-coordinate is the average accuracy difference between the top-unit-ablated and unabl… view at source ↗
Figure 20
Figure 20. Figure 20: Visualization of the results of the ablation [PITH_FULL_IMAGE:figures/full_fig_p018_20.png] view at source ↗
Figure 22
Figure 22. Figure 22: Results of the 2-fold cross-validation analy [PITH_FULL_IMAGE:figures/full_fig_p019_22.png] view at source ↗
read the original abstract

Large language models (LLMs) can reliably distinguish grammatical from ungrammatical sentences, but how grammatical knowledge is represented within the models remains an open question. We investigate whether different syntactic phenomena recruit shared or distinct components in LLMs. Using a functional localization approach inspired by cognitive neuroscience, we identify the LLM units most responsive to 67 English syntactic phenomena in seven open-weight models. These units are consistently recruited across sentences containing the phenomena and causally support the models' syntactic performance. Critically, different types of syntactic agreement (e.g., subject-verb, anaphor, determiner-noun) recruit overlapping sets of units, suggesting that agreement constitutes a meaningful functional category for LLMs. This pattern holds in English, Russian, and Chinese; and further, in a cross-lingual analysis of 57 diverse languages, structurally more similar languages share more units for subject-verb agreement. Taken together, these findings reveal that syntactic agreement-a critical marker of syntactic dependencies-constitutes a meaningful category within LLMs' representational spaces.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents an investigation into how different syntactic phenomena, particularly various types of agreement, are represented in large language models (LLMs). Using a functional localization method inspired by neuroscience, the authors identify units in seven open-weight LLMs that respond most strongly to 67 English syntactic phenomena. They demonstrate that these units are consistently activated by relevant sentences and causally influence the models' ability to handle syntactic tasks. A key finding is that different agreement types—such as subject-verb, anaphor, and determiner-noun—share overlapping units, indicating that agreement may form a distinct functional category in LLMs' internal representations. This pattern is shown to generalize to Russian and Chinese, and cross-lingual analysis across 57 languages reveals greater unit sharing for subject-verb agreement between structurally similar languages.

Significance. If the central findings are robust to controls for stimulus confounds, this work would make a notable contribution to the field of LLM interpretability by providing evidence that syntactic agreement is encoded as a coherent category rather than through disparate mechanisms. The neuroscience-inspired functional localization approach, combined with causal interventions and cross-linguistic comparisons, offers a promising framework for understanding structured knowledge in neural networks. It could influence research on model alignment with human-like grammatical processing and cross-lingual generalization.

major comments (2)
  1. [Methods] Methods (functional localization): The procedure selects units most responsive to sentences containing each syntactic phenomenon, but provides no explicit controls or matching for confounding variables such as sentence length, lexical predictability, or overall syntactic complexity across stimulus sets for different agreement types. If these differ systematically, the reported overlap for subject-verb, anaphor, and determiner-noun agreement could arise from shared non-syntactic features rather than a common functional category for agreement. This is load-bearing for the central claim that agreement constitutes a meaningful category.
  2. [Results] Results (causal interventions): The abstract asserts that the identified units causally support syntactic performance, yet the exact intervention technique (e.g., activation patching, ablation) and any specificity controls to distinguish agreement processing from general syntactic or semantic effects are not detailed. This leaves open whether the interventions confirm agreement-specific involvement or broader sentence-level functions.
minor comments (2)
  1. [Abstract] Abstract: The reference to '67 English syntactic phenomena' would benefit from a brief indication of selection criteria or a pointer to the supplementary materials for the full list and categorization.
  2. [Cross-lingual analysis] Cross-lingual analysis: Clarify how structural similarity between languages was quantified (e.g., specific metrics or treebank features) to support the claim of greater unit sharing.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We agree that clarifying controls for potential confounds and providing more specifics on causal interventions will strengthen the manuscript. We address each major comment below and will incorporate revisions to improve transparency and robustness.

read point-by-point responses

  1. Referee: [Methods] Methods (functional localization): The procedure selects units most responsive to sentences containing each syntactic phenomenon, but provides no explicit controls or matching for confounding variables such as sentence length, lexical predictability, or overall syntactic complexity across stimulus sets for different agreement types. If these differ systematically, the reported overlap for subject-verb, anaphor, and determiner-noun agreement could arise from shared non-syntactic features rather than a common functional category for agreement. This is load-bearing for the central claim that agreement constitutes a meaningful category.

    Authors: We acknowledge the validity of this concern. Our stimulus construction aimed to isolate syntactic phenomena by using minimal pairs where possible, but we did not perform explicit matching or regression controls for sentence length, lexical predictability, or overall complexity across the full set of 67 phenomena. To address this directly, we will add supplementary analyses in the revised manuscript that include length-matched controls, perplexity-based predictability measures, and partial correlation or regression models to isolate syntactic effects. These additions will test whether the observed unit overlap for agreement types persists after accounting for non-syntactic factors. revision: yes

  2. Referee: [Results] Results (causal interventions): The abstract asserts that the identified units causally support syntactic performance, yet the exact intervention technique (e.g., activation patching, ablation) and any specificity controls to distinguish agreement processing from general syntactic or semantic effects are not detailed. This leaves open whether the interventions confirm agreement-specific involvement or broader sentence-level functions.

    Authors: We appreciate this observation. The manuscript describes causal interventions via targeted activation patching on the localized units, showing performance drops on agreement-related tasks. However, we agree that greater detail on the precise technique and specificity controls (e.g., comparisons to general syntactic or semantic interventions) is warranted for clarity. In the revision, we will expand the methods and results sections to fully specify the patching procedure, include control interventions on non-agreement units, and report quantitative comparisons demonstrating that effects are stronger for agreement tasks than for matched general syntactic or semantic benchmarks. revision: yes

Circularity Check

0 steps flagged

Empirical localization and overlap analysis without circular reduction

full rationale

The paper derives its claims from direct empirical procedures: measuring unit activations in response to sentences containing specific syntactic phenomena, selecting the most responsive units, and testing their causal role via interventions. Overlap among units for different agreement types is reported as an observed pattern across models and languages, not as a quantity forced by definition, a fitted parameter, or a self-citation chain. No equations or derivations reduce the central result to its inputs by construction; the functional localization is applied to open models using stimulus sets whose properties are external to the analysis itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim rests primarily on the validity of the neuroscience-inspired localization method and the interpretation that overlapping responsive units indicate a functional category; no new entities are postulated and no free parameters are explicitly fitted in the abstract.

axioms (1)
  • domain assumption Units identified as most responsive via the localization procedure are causally involved in syntactic processing rather than reflecting correlated but non-syntactic features.
    This assumption underpins both the identification of units and the causal claims about syntactic performance.

pith-pipeline@v0.9.0 · 5487 in / 1315 out tokens · 86653 ms · 2026-05-17T02:38:15.216124+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

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

  1. Fine-Grained Analysis of Shared Syntactic Mechanisms in Language Models

    cs.CL 2026-04 unverdicted novelty 6.0

    Language models employ a highly localized shared mechanism for filler-gap dependencies but no unified mechanism for NPI licensing, and activation patching generalizes better than supervised alignment search.

Reference graph

Works this paper leans on

69 extracted references · 69 canonical work pages · cited by 1 Pith paper · 7 internal anchors

  1. [1]

    Badr AlKhamissi, Greta Tuckute, Antoine Bosselut, and Martin Schrimpf. 2025. https://doi.org/10.18653/v1/2025.naacl-long.544 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 Linguistics: Human...

    work page doi:10.18653/v1/2025.naacl-long.544 2025

  2. [2]

    Aixiu An, Peng Qian, Ethan Wilcox, and Roger Levy. 2019. https://arxiv.org/abs/1909.04625 Representation of constituents in neural language models: Coordination phrase as a case study . arXiv preprint arXiv:1909.04625

    work page arXiv 2019

  3. [3]

    Chang, James A

    Catherine Arnett, Tyler A. Chang, James A. Michaelov, and Benjamin K. Bergen. 2025. https://arxiv.org/abs/2503.03962 On the acquisition of shared grammatical representations in bilingual language models . arXiv preprint arXiv:2503.03962

    work page arXiv 2025

  4. [4]

    Ezgi Başar, Francesca Padovani, Jaap Jumelet, and Arianna Bisazza. 2025. https://arxiv.org/abs/2506.13487 Tur BL i MP : A T urkish benchmark of linguistic minimal pairs . arXiv preprint arXiv:2506.13487

    work page arXiv 2025

  5. [5]

    Yonatan Belinkov, Llu \' s M \`a rquez, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and James Glass. 2017. Evaluating layers of representation in neural machine translation on part-of-speech and semantic tagging tasks. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1--10

    work page 2017

  6. [6]

    Sasha Boguraev, Christopher Potts, and Kyle Mahowald. 2025. https://arxiv.org/abs/2505.16002 Causal interventions reveal shared structure across E nglish filler-gap constructions . arXiv preprint arXiv:2505.16002

    work page arXiv 2025

  7. [7]

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

    work page arXiv 2025

  8. [8]

    Lyle Campbell and Ver \'o nica Grondona. 2008. Ethnologue: Languages of the world. Language, 84(3):636--641

    work page 2008

  9. [9]

    Chi, John Hewitt, and Christopher D

    Ethan A. Chi, John Hewitt, and Christopher D. Manning. 2020. https://doi.org/10.18653/v1/2020.acl-main.493 Finding universal grammatical relations in multilingual BERT . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5564--5577, Online. Association for Computational Linguistics

    work page doi:10.18653/v1/2020.acl-main.493 2020

  10. [10]

    Noam Chomsky. 1981. https://doi.org/doi:10.1515/tlir.1981.1.1.3 On the representation of form and function . The Linguistic Review, 1(1):3--40

    work page doi:10.1515/tlir.1981.1.1.3 1981

  11. [11]

    Chris Collins and Richard Kayne. 2011. Syntactic structures of the world’s languages

    work page 2011

  12. [12]

    Andrea Gregor de Varda and Marco Marelli. 2023. Data-driven cross-lingual syntax: An agreement study with massively multilingual models. Computational Linguistics, 49(2):261--299

    work page 2023

  13. [13]

    Google DeepMind. 2025. https://arxiv.org/pdf/2503.19786 Gemma 3 technical report . arXiv preprint arXiv:2503.19786

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    Deep S eek. 2024. https://arxiv.org/abs/2401.02954 Deep S eek LLM : Scaling open-source language models with longtermism . arXiv preprint arXiv:2401.02954

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  15. [15]

    Dryer and Martin Haspelmath, editors

    Matthew S. Dryer and Martin Haspelmath, editors. 2013. https://doi.org/10.5281/zenodo.13950591 WALS Online (v2020.4) . Zenodo

    work page doi:10.5281/zenodo.13950591 2013

  16. [16]

    Xufeng Duan, Zhaoqian Yao, Yunhao Zhang, Shaonan Wang, and Zhenguang G. Cai. 2025. https://arxiv.org/abs/2505.19548 How syntax specialization emerges in language models . arXiv preprint arXiv:2505.19548

    work page arXiv 2025

  17. [17]

    Nadir Durrani, Hassan Sajjad, Fahim Dalvi, and Yonatan Belinkov. 2020. https://arxiv.org/abs/2010.02695 Analyzing individual neurons in pre-trained language models . arXiv preprint arXiv:2010.02695

    work page arXiv 2020

  18. [18]

    Evelina Fedorenko, Po-Jang Hsieh, Alfonso Nieto-Casta \ n \'o n, Susan Whitfield-Gabrieli, and Nancy Kanwisher. 2010. New method for f MRI investigations of language: defining ROI s functionally in individual subjects. Journal of neurophysiology, 104(2):1177--1194

    work page 2010

  19. [19]

    Costa-juss \`a

    Javier Ferrando and Marta R. Costa-juss \`a . 2024. On the similarity of circuits across languages: a case study on the subject-verb agreement task. In ICML 2024 Workshop on Mechanistic Interpretability

    work page 2024

  20. [20]

    Richard Futrell, Ethan Wilcox, Takashi Morita, Peng Qian, Miguel Ballesteros, and Roger Levy. 2019. Neural language models as psycholinguistic subjects: Representations of syntactic state. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Sh...

    work page 2019

  21. [21]

    Jon Gauthier, Jennifer Hu, Ethan Wilcox, Peng Qian, and Roger Levy. 2020. https://doi.org/10.18653/v1/2020.acl-demos.10 S yntax G ym: 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 Comput...

    work page doi:10.18653/v1/2020.acl-demos.10 2020

  22. [22]

    Atticus Geiger, Zhengxuan Wu, Christopher Potts, Thomas Icard, and Noah Goodman. 2024. Finding alignments between interpretable causal variables and distributed neural representations. In Causal Learning and Reasoning, pages 160--187. PMLR

    work page 2024

  23. [23]

    Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. https://doi.org/10.18653/v1/N18-1108 Colorless green recurrent networks dream hierarchically . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) , ...

    work page doi:10.18653/v1/n18-1108 2018

  24. [24]

    Bobby He, Lorenzo Noci, Daniele Paliotta, Imanol Schlag, and Thomas Hofmann. 2024. Understanding and minimising outlier features in transformer training. Advances in Neural Information Processing Systems, 37:83786--83846

    work page 2024

  25. [25]

    John Hewitt and Christopher D. Manning. 2019. https://doi.org/10.18653/v1/N19-1419 A structural probe for finding syntax in word representations . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages 4129--4138, Minneapol...

    work page doi:10.18653/v1/n19-1419 2019

  26. [26]

    Huang, Kuan-Hao Huang, and Kai-Wei Chang

    James Y. Huang, Kuan-Hao Huang, and Kai-Wei Chang. 2021. https://arxiv.org/abs/2104.05115 Disentangling semantics and syntax in sentence embeddings with pre-trained language models . arXiv preprint arXiv:2104.05115

    work page arXiv 2021

  27. [27]

    Vo, Leila Wehbe, and Alexander G

    Shailee Jain, Vy A. Vo, Leila Wehbe, and Alexander G. Huth. 2024. Computational language modeling and the promise of in silico experimentation. Neurobiology of Language, 5(1):80--106

    work page 2024

  28. [28]

    Ganesh Jawahar, Beno \^i t Sagot, and Djam \'e Seddah. 2019. https://doi.org/10.18653/v1/P19-1356 What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651--3657, Florence, Italy. Association for Computational Linguistics

    work page doi:10.18653/v1/p19-1356 2019

  29. [29]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. https://arxiv.org/abs/2310.0...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [30]

    Cameron Jones and Ben Bergen. 2024. Does GPT -4 pass the T uring test? In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5183--5210

    work page 2024

  31. [31]

    Jones and Benjamin K

    Cameron R. Jones and Benjamin K. Bergen. 2025. https://arxiv.org/abs/2503.23674 Large language models pass the T uring test . arXiv preprint arXiv:2503.23674

    work page arXiv 2025

  32. [32]

    Jaap Jumelet, Leonie Weissweiler, and Arianna Bisazza. 2025. https://arxiv.org/abs/2504.02768 Multi BL i MP 1.0: A massively multilingual benchmark of linguistic minimal pairs . arXiv preprint arXiv:2504.02768

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    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

    work page 2021

  34. [34]

    Yair Lakretz, Germ \'a n Kruszewski, Th \'e o Desbordes, Dieuwke Hupkes, Stanislas Dehaene, and Marco Baroni. 2019. The emergence of number and syntax units in LSTM language models. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Pap...

    work page 2019

  35. [35]

    Lepori, Thomas Serre, and Ellie Pavlick

    Michael A. Lepori, Thomas Serre, and Ellie Pavlick. 2023. Uncovering intermediate variables in transformers using circuit probing. In First Conference on Language Modeling

    work page 2023

  36. [36]

    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

    work page doi:10.1162/tacl_a_00115 2016

  37. [37]

    Mortensen, Ke Lin, Katherine Kairis, Carlisle Turner, and Lori Levin

    Patrick Littell, David R. Mortensen, Ke Lin, Katherine Kairis, Carlisle Turner, and Lori Levin. 2017. Uriel and lang2vec: Representing languages as typological, geographical, and phylogenetic vectors. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 8--14

    work page 2017

  38. [38]

    Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models

    Samuel Marks, Can Rager, Eric J. Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. 2024. https://arxiv.org/abs/2403.19647 Sparse feature circuits: Discovering and editing interpretable causal graphs in language models . arXiv preprint arXiv:2403.19647

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    Rebecca Marvin and Tal Linzen. 2018. Targeted syntactic evaluation of language models. In 2018 Conference on Empirical Methods in Natural Language Processing, EMNLP 2018, pages 1192--1202. Association for Computational Linguistics

    work page 2018

  40. [40]

    Thomas McCoy, Ellie Pavlick, and Tal Linzen

    R. Thomas McCoy, Ellie Pavlick, and Tal Linzen. 2020. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In 57th Annual Meeting of the Association for Computational Linguistics, ACL 2019, pages 3428--3448. Association for Computational Linguistics (ACL)

    work page 2020

  41. [41]

    Meta. 2024. https://huggingface.co/blog/llama32 Llama can now see and run on your device - welcome L lama 3.2

    work page 2024

  42. [42]

    James Michaelov, Catherine Arnett, Tyler Chang, and Ben Bergen. 2023. Structural priming demonstrates abstract grammatical representations in multilingual language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3703--3720

    work page 2023

  43. [43]

    Microsoft. 2025. https://arxiv.org/abs/2503.01743 Phi-4-mini technical report: Compact yet powerful multimodal language models via M ixture-of- L o RA s . arXiv preprint arXiv:2503.01743

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    Kanishka Misra. 2022. https://arxiv.org/abs/2203.13112 minicons: Enabling flexible behavioral and representational analyses of transformer language models . arXiv preprint arXiv:2203.13112

    work page arXiv 2022

  45. [45]

    Aaron Mueller, Jannik Brinkmann, Millicent Li, Samuel Marks, Koyena Pal, Nikhil Prakash, Can Rager, Aruna Sankaranarayanan, Arnab Sen Sharma, Jiuding Sun, Eric Todd, David Bau, and Yonatan Belinkov. 2024. https://arxiv.org/abs/2408.01416 The quest for the right mediator: A history, survey, and theoretical grounding of causal interpretability . arXiv prepr...

    work page arXiv 2024

  46. [46]

    Aaron Mueller, Yu Xia, and Tal Linzen. 2022. https://doi.org/10.18653/v1/2022.conll-1.8 Causal analysis of syntactic agreement neurons in multilingual language models . In Proceedings of the 26th Conference on Computational Natural Language Learning (CoNLL), pages 95--109, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics

    work page doi:10.18653/v1/2022.conll-1.8 2022

  47. [47]

    Chi, Richard Futrell, and Kyle Mahowald

    Isabel Papadimitriou, Ethan A. Chi, Richard Futrell, and Kyle Mahowald. 2021. https://arxiv.org/abs/2101.11043 Deep subjecthood: Higher-order grammatical features in multilingual BERT . arXiv preprint arXiv:2101.11043

    work page arXiv 2021

  48. [48]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners

    work page 2019

  49. [49]

    Daking Rai, Yilun Zhou, Shi Feng, Abulhair Saparov, and Ziyu Yao. 2024. https://arxiv.org/abs/2407.02646 A practical review of mechanistic interpretability for transformer-based language models . arXiv preprint arXiv:2407.02646

    work page arXiv 2024

  50. [50]

    Aniketh Janardhan Reddy and Leila Wehbe. 2021. https://proceedings.neurips.cc/paper_files/paper/2021/file/51a472c08e21aef54ed749806e3e6490-Paper.pdf Can f MRI reveal the representation of syntactic structure in the brain? In Advances in Neural Information Processing Systems, volume 34, pages 9843--9856. Curran Associates, Inc

    work page 2021

  51. [51]

    Rebecca Saxe, Matthew Brett, and Nancy Kanwisher. 2006. Divide and conquer: a defense of functional localizers. Neuroimage, 30(4):1088--1096

    work page 2006

  52. [52]

    Yixiao Song, Kalpesh Krishna, Rajesh Bhatt, and Mohit Iyyer. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.305 SLING : S ino linguistic evaluation of large language models . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4606--4634, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics

    work page doi:10.18653/v1/2022.emnlp-main.305 2022

  53. [53]

    Karolina Stanczak, Edoardo Ponti, Lucas Torroba Hennigen, Ryan Cotterell, and Isabelle Augenstein. 2022. https://doi.org/10.18653/v1/2022.naacl-main.114 Same neurons, different languages: Probing morphosyntax in multilingual pre-trained models . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Lingui...

    work page doi:10.18653/v1/2022.naacl-main.114 2022

  54. [54]

    Giulio Starace, Konstantinos Papakostas, Rochelle Choenni, Apostolos Panagiotopoulos, Matteo Rosati, Alina Leidinger, and Ekaterina Shutova. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.476 Probing LLM s for joint encoding of linguistic categories . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7158--7179, Singap...

    work page doi:10.18653/v1/2023.findings-emnlp.476 2023

  55. [55]

    Michelle Suijkerbuijk, Zoë Prins, Marianne de Heer Kloots, Willem Zuidema, and Stefan L. Frank. 2025. https://doi.org/10.1162/coli_a_00559 BL i MP-NL : A corpus of D utch minimal pairs and acceptability judgments for language model evaluation . Computational Linguistics, pages 1--35

    work page doi:10.1162/coli_a_00559 2025

  56. [56]

    Ekaterina Taktasheva, Maxim Bazhukov, Kirill Koncha, Alena Fenogenova, Ekaterina Artemova, and Vladislav Mikhailov. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.522 R u BL i MP : R ussian benchmark of linguistic minimal pairs . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 9268--9299, Miami, Florida,...

    work page doi:10.18653/v1/2024.emnlp-main.522 2024

  57. [57]

    Tianyi Tang, Wenyang Luo, Haoyang Huang, Dongdong Zhang, Xiaolei Wang, Wayne Xin Zhao, Furu Wei, and Ji-Rong Wen. 2024. Language-specific neurons: The key to multilingual capabilities in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5701--5715

    work page 2024

  58. [58]

    Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. https://doi.org/10.18653/v1/P19-1452 BERT rediscovers the classical NLP pipeline . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593--4601, Florence, Italy. Association for Computational Linguistics

    work page doi:10.18653/v1/p19-1452 2019

  59. [59]

    Technology Innovation Institute (TII). 2024. https://huggingface.co/blog/falcon3 Welcome to the F alcon 3 family of open models!

    work page 2024

  60. [60]

    Greta Tuckute, Nancy Kanwisher, and Evelina Fedorenko. 2024. Language in brains, minds, and machines. Annual Review of Neuroscience, 47(2024):277--301

    work page 2024

  61. [61]

    o tgen, and Hinrich Sch \

    Mingyang Wang, Heike Adel, Lukas Lange, Yihong Liu, Ercong Nie, Jannik Str \"o tgen, and Hinrich Sch \"u tze. 2025. https://arxiv.org/abs/2504.04264 Lost in multilinguality: Dissecting cross-lingual factual inconsistency in transformer language models . arXiv preprint arXiv:2504.04264

    work page arXiv 2025

  62. [62]

    Alex Warstadt, Yu Cao, Ioana Grosu, Wei Peng, Hagen Blix, Yining Nie, Anna Alsop, Shikha Bordia, Haokun Liu, Alicia Parrish, Sheng-Fu Wang, Jason Phang, Anhad Mohananey, Phu Mon Htut, Paloma Jeretic, and Samuel R. Bowman. 2019. https://arxiv.org/abs/1909.02597 Investigating BERT 's knowledge of language: Five analysis methods with NPI s . arXiv preprint a...

    work page arXiv 2019

  63. [63]

    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 Linguistics, 8:377--392

    work page doi:10.1162/tacl_a_00321 2020

  64. [64]

    Ethan Wilcox, Roger Levy, Takashi Morita, and Richard Futrell. 2018. What do RNN language models learn about filler-gap dependencies? In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 211--221

    work page 2018

  65. [65]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, and 1 others. 2019. https://arxiv.org/abs/1910.03771 Hugging F ace's transformers: State-of-the-art natural language processing . arXiv preprint arXiv:1910.03771

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  66. [66]

    Zhengxuan Wu, Atticus Geiger, Thomas Icard, Christopher Potts, and Noah Goodman. 2023. Interpretability at scale: Identifying causal mechanisms in A lpaca. Advances in neural information processing systems, 36:78205--78226

    work page 2023

  67. [67]

    Xiaohan Zhang, Shaonan Wang, Nan Lin, Jiajun Zhang, and Chengqing Zong. 2022. https://doi.org/10.1609/aaai.v36i10.21427 Probing word syntactic representations in the brain by a feature elimination method . Proceedings of the AAAI Conference on Artificial Intelligence, 36(10):11721--11729

    work page doi:10.1609/aaai.v36i10.21427 2022

  68. [68]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

    work page

  69. [69]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page