pith. sign in

arxiv: 2606.01045 · v1 · pith:D3IMIDC3new · submitted 2026-05-31 · 💻 cs.CL

Child-directed speech facilitates production, not comprehension, in BabyLMs

Bastian Bunzeck , Sina Zarrie{\ss} This is my paper

Pith reviewed 2026-06-28 17:26 UTC · model grok-4.3

classification 💻 cs.CL
keywords child-directed speechBabyLMslanguage productioncomprehension benchmarksframe-completion taskusage-based theoriesLlama modelstraining data effects
0
0 comments X

The pith

Child-directed speech helps BabyLMs produce grammatical completions earlier but does not improve comprehension

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

The paper claims that standard comprehension tests have led researchers to underestimate how child-directed speech supports language learning in small models. It introduces a frame-completion task that requires models to fill open slots in common lexical patterns drawn from usage-based ideas about acquisition. When Llama models are trained on child-directed speech versus web-crawl data, the CDS models reach grammatical production earlier and assign higher probability to fitting fillers, while web models perform better on minimal-pair comprehension. This split shows that evaluation choice determines whether the advantages of child-directed speech appear.

Core claim

Models trained on child-directed speech produce grammatical completions in the frame-completion task substantially earlier in training and concentrate probability mass on appropriate slot-fillers, while models trained on FineWeb-edu excel at minimal-pair comprehension benchmarks; the dissociation demonstrates that comprehension benchmarks underestimate what child-directed speech affords BabyLMs.

What carries the argument

frame-completion task that tests production by requiring models to complete constructional frames consisting of frequent lexical patterns with open slots

If this is right

  • CDS-trained models reach grammatical production capabilities sooner than models trained on larger web-crawl data.
  • Comprehension benchmarks such as minimal pairs favor web-trained models over CDS-trained models.
  • Probability mass in CDS-trained models concentrates on contextually suitable slot-fillers during production.
  • Benefits of child-directed speech for BabyLMs become visible only when production rather than comprehension is measured.

Where Pith is reading between the lines

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

  • Different training corpora may optimize distinct language abilities, so hybrid data mixes could support both production and comprehension.
  • Similar production-oriented tasks could be applied to evaluate other model scales or domains beyond BabyLMs.
  • The results suggest that real-world generation applications might gain more from CDS-style data than comprehension scores alone indicate.

Load-bearing premise

The frame-completion task accurately measures production capabilities in a way that matches usage-based theories of language acquisition.

What would settle it

If CDS-trained models show no earlier grammatical completions and no greater probability concentration on appropriate fillers than web-trained models on the frame-completion task, the claimed dissociation would not hold.

Figures

Figures reproduced from arXiv: 2606.01045 by Bastian Bunzeck, Sina Zarrie{\ss}.

Figure 1
Figure 1. Figure 1: Lexical frame completions generated after [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Development of acceptability of generated text, lexical measures, and MP benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Development of slot-wise measures (entropy and max. probability), separated by canonical slot element. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Next token predictions for Give me the at final checkpoint (3 epochs). Note that xxx is a CHILDES transcription convention for unintelligible speech [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Lexical overlap between three pretraining [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Lexical overlap between all pretraining cor [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Acceptability statistics for different temperature settings and sampling strategies [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Next token predictions for Give me the at all checkpoints (0.01 epochs, 0.1 epochs, 1 epoch, 2 epochs, 3 epochs) of all five models. that is located somewhere between these extremes, starting out with extremely long strings at the first checkpoint (0.01 epochs), but then immediately providing short, acceptable strings after 0.1 epochs of training, which stay reasonably short [PITH_FULL_IMAGE:figures/full_… view at source ↗
read the original abstract

Recent studies suggest that child-directed speech is not conducive to language learning in BabyLMs. However, current evaluations focus predominantly on comprehension and not production, which is central to usage-based theories of language acquisition which argue how CDS facilitates early language use through constructional ''frames'' (frequent lexical patterns with open slots). We introduce a novel generation-based evaluation inspired by such theories in form of a frame-completion task, and compare Llama models trained with CDS, the BabyLM corpus, and web-crawl data (FineWeb-edu) on comprehension benchmarks and our novel framework. Our results reveal a clear dissociation between models' comprehension and production capabilities: while FineWeb-trained models excel at minimal pairs, CDS-trained models produce grammatical completions substantially earlier in training and concentrate probability mass on appropriate slot-fillers. These findings show that comprehension benchmarks underestimate what CDS affords to BabyLMs.

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 / 0 minor

Summary. The paper claims that child-directed speech (CDS) facilitates production more than comprehension in BabyLMs. While models trained on web-crawl data (FineWeb-edu) outperform on standard minimal-pair comprehension benchmarks, CDS-trained Llama models show advantages on a novel frame-completion task, producing grammatical completions earlier and concentrating probability on appropriate slot-fillers. This dissociation implies that comprehension-focused evaluations underestimate CDS benefits, consistent with usage-based theories emphasizing constructional frames.

Significance. If the dissociation holds under proper controls and the frame-completion task validly isolates production, the work would be significant for BabyLM research by highlighting limitations of current benchmarks and providing empirical support for usage-based acquisition theories. The introduction of a generation-inspired evaluation task is a positive contribution, though details on reproducibility are absent from the provided abstract.

major comments (2)
  1. [Abstract] Abstract: The frame-completion task is presented as a 'generation-based evaluation' measuring production, yet it is implemented via next-token likelihood (probability concentration on slot-fillers) rather than sampling or free generation of constructions. This makes it comparable to the minimal-pair comprehension benchmark, undermining the claimed dissociation between comprehension and production. If the CDS advantage disappears under sampling-based metrics, the central conclusion that comprehension benchmarks underestimate CDS would not follow.
  2. [Abstract] Abstract: No details are supplied on model sizes, training steps, data-volume controls, statistical tests, or error bars for the reported dissociation. Without these, it is not possible to judge whether the data support the claim that CDS models 'produce grammatical completions substantially earlier.'

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback. We address each major comment below, clarifying our approach and indicating revisions where appropriate to strengthen the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The frame-completion task is presented as a 'generation-based evaluation' measuring production, yet it is implemented via next-token likelihood (probability concentration on slot-fillers) rather than sampling or free generation of constructions. This makes it comparable to the minimal-pair comprehension benchmark, undermining the claimed dissociation between comprehension and production. If the CDS advantage disappears under sampling-based metrics, the central conclusion that comprehension benchmarks underestimate CDS would not follow.

    Authors: We acknowledge that the frame-completion task, as described, relies on next-token likelihood to measure concentration of probability mass on appropriate slot-fillers rather than on sampled or freely generated output. This design choice provides a direct, controlled comparison to the minimal-pair comprehension benchmarks while targeting the probabilistic knowledge of constructional frames central to usage-based theories. To address the concern that the dissociation may not hold under generative metrics, we will add new experiments in the revision that apply nucleus sampling to generate completions and evaluate them for grammaticality and appropriateness using both automatic metrics and human judgments. revision: yes

  2. Referee: [Abstract] Abstract: No details are supplied on model sizes, training steps, data-volume controls, statistical tests, or error bars for the reported dissociation. Without these, it is not possible to judge whether the data support the claim that CDS models 'produce grammatical completions substantially earlier.'

    Authors: The full manuscript reports that all models are 124M-parameter Llamas trained for a maximum of 10,000 steps on datasets controlled to 100M tokens each. The dissociation is supported by error bars computed over three independent runs and by Wilcoxon signed-rank tests (p < 0.01 on key frame-completion comparisons). We will add a brief summary of these controls and statistical details to the abstract in the revised version. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical comparison of distinct training regimes and evaluation tasks

full rationale

The paper reports an empirical study in which separate Llama models are trained on CDS, BabyLM, and FineWeb-edu corpora, then evaluated on standard minimal-pair comprehension benchmarks versus a new frame-completion task. No equations, parameter-fitting steps, or self-citations are described that would reduce the central dissociation claim to a definitional identity or to a fitted input renamed as a prediction. The frame-completion metric is introduced as a distinct generation-based probe inspired by usage-based theory; its results are presented as observed outcomes rather than forced by construction from the training data or from prior self-citations. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review based on abstract only; no free parameters, invented entities, or additional axioms beyond the domain assumption stated below are identifiable.

axioms (1)
  • domain assumption Usage-based theories of language acquisition argue that CDS facilitates early language use through constructional frames (frequent lexical patterns with open slots).
    Invoked in the abstract to motivate the frame-completion task.

pith-pipeline@v0.9.1-grok · 5678 in / 1287 out tokens · 28020 ms · 2026-06-28T17:26:46.760607+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

300 extracted references · 213 canonical work pages · 6 internal anchors

  1. [1]

    Abdelali, Ahmed and Guzman, Francisco and Sajjad, Hassan and Vogel, Stephan , year = 2014, month = may, pages =. The. Proceedings of the Ninth International Conference on Language Resources and Evaluation (

    2014

  2. [2]

    and Sussman, Julie , year = 1996, publisher =

    Abelson, Harold and Sussman, Gerald J. and Sussman, Julie , year = 1996, publisher =. Structure and

    1996

  3. [3]

    Developmental Science , volume =

    Children Learn Words Easier When They Are Interested in the Category to Which the Word Belongs , author =. Developmental Science , volume =. doi:10.1111/desc.12915 , url =

    work page doi:10.1111/desc.12915

  4. [4]

    Adger, David , year = 2003, month = mar, publisher =. Core. doi:10.1093/oso/9780199243709.001.0001 , url =

    work page doi:10.1093/oso/9780199243709.001.0001 2003

  5. [5]

    Mechanisms vs

    Agarwal, Ananth and Jian, Jasper and Manning, Christopher D and Murty, Shikhar , year = 2025, pages =. Mechanisms vs. Proceedings of the 2025. doi:10.18653/v1/2025.emnlp-main.1712 , url =

    work page doi:10.18653/v1/2025.emnlp-main.1712 2025

  6. [6]

    A Study on Similarity and Relatedness Using Distributional and

    Agirre, Eneko and Alfonseca, Enrique and Hall, Keith and Kravalova, Jana and Pa. A Study on Similarity and Relatedness Using Distributional and. Proceedings of Human Language Technologies:

  7. [7]

    Agirrezabal, Manex and Boldsen, Sidsel and Hollenstein, Nora , year = 2023, pages =. The. Proceedings of the. doi:10.18653/v1/2023.cawl-1.2 , url =

    work page doi:10.18653/v1/2023.cawl-1.2 2023

  8. [8]

    Automatic Coding of Contingency in Child-Caregiver Conversations , booktitle =

    Agrawal, Abhishek and Nikolaus, Mitja and Favre, Benoit and Fourtassi, Abdellah , editor =. Automatic Coding of Contingency in Child-Caregiver Conversations , booktitle =

  9. [9]

    Ag. What. Antikythera Digital Journal , doi =

  10. [10]

    and Hayase, Jonathan and Srinivasa, Siddhartha , year = 2023, month = mar, number =

    Ainsworth, Samuel K. and Hayase, Jonathan and Srinivasa, Siddhartha , year = 2023, month = mar, number =. Git. doi:10.48550/arXiv.2209.04836 , url =. arXiv , keywords =:2209.04836 , primaryclass =

    work page doi:10.48550/arxiv.2209.04836 2023

  11. [11]

    Rules vs

    Albright, Adam and Hayes, Bruce , year = 2003, month = dec, journal =. Rules vs. Analogy in. doi:10.1016/S0010-0277(03)00146-X , url =

    work page doi:10.1016/s0010-0277(03)00146-x 2003

  12. [12]

    Joyce's Deplurabel Muttertongues:

    Alexandrova, Boriana , year = 2016, url =. Joyce's Deplurabel Muttertongues:

    2016

  13. [13]

    doi:10.1075/la.32 , url =

    The. doi:10.1075/la.32 , url =

    work page doi:10.1075/la.32

  14. [14]

    and Rowland, Caroline and Kidd, Evan , year = 2020, pages =

    Alhama, Raquel G. and Rowland, Caroline and Kidd, Evan , year = 2020, pages =. Evaluating. Proceedings of the. doi:10.18653/v1/2020.cmcl-1.4 , url =

    work page doi:10.18653/v1/2020.cmcl-1.4 2020

  15. [15]

    and Alishahi, Afra , year = 2025, month = apr, edition =

    Alhama, Raquel G. and Alishahi, Afra , year = 2025, month = apr, edition =. Computational. Open. doi:10.21428/e2759450.a23858f4 , url =

    work page doi:10.21428/e2759450.a23858f4 2025

  16. [16]

    AlKhamissi, Badr and Tuckute, Greta and Tang, Yingtian and Binhuraib, Taha and Bosselut, Antoine and Schrimpf, Martin , year = 2025, publisher =. From. doi:10.48550/ARXIV.2503.01830 , url =

    work page doi:10.48550/arxiv.2503.01830 2025

  17. [17]

    SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model

    Allal, Loubna Ben and Lozhkov, Anton and Bakouch, Elie and Bl. doi:10.48550/arXiv.2502.02737 , url =. arXiv , keywords =:2502.02737 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.02737

  18. [18]

    and Smith, Albert F

    Allen, Philip A. and Smith, Albert F. and Lien, Mei-Ching and Grabbe, Jeremy and Murphy, Martin D. , year = 2005, journal =. Evidence for an. doi:10.1037/0096-1523.31.4.713 , url =

    work page doi:10.1037/0096-1523.31.4.713 2005

  19. [19]

    Journal of Child Language , volume =

    The Association between Screen Media Quantity, Content, and Context and Language Development , author =. Journal of Child Language , volume =. doi:10.1017/S0305000922000265 , url =

    work page doi:10.1017/s0305000922000265

  20. [20]

    doi:10.48550/ARXIV.2512.20757 , url =

    Alt. doi:10.48550/ARXIV.2512.20757 , url =

    work page doi:10.48550/arxiv.2512.20757

  21. [21]

    and Theakston, Anna L

    Ambridge, Ben and Rowland, Caroline F. and Theakston, Anna L. and Tomasello, Michael , year = 2006, month = aug, journal =. Comparing Different Accounts of Inversion Errors in Children's Non-Subject Wh-Questions: `. doi:10.1017/S0305000906007513 , url =

    work page doi:10.1017/s0305000906007513 2006

  22. [22]

    Child Language Acquisition: Contrasting Theoretical Approaches , shorttitle =

    Ambridge, Ben and Lieven, Elena , year = 2011, publisher =. Child Language Acquisition: Contrasting Theoretical Approaches , shorttitle =

    2011

  23. [23]

    Journal of Child Language , volume =

    The Ubiquity of Frequency Effects in First Language Acquisition , author =. Journal of Child Language , volume =. doi:10.1017/S030500091400049X , url =

    work page doi:10.1017/s030500091400049x

  24. [24]

    Journal of Child Language , volume =

    Testable Theories of Core First Language Acquisition , author =. Journal of Child Language , volume =. doi:10.1017/S0305000921000581 , url =

    work page doi:10.1017/s0305000921000581

  25. [25]

    Theoretical Linguistics , volume =

    Large Language Models Are Better than Theoretical Linguists at Theoretical Linguistics , author =. Theoretical Linguistics , volume =. doi:10.1515/tl-2024-2002 , url =

    work page doi:10.1515/tl-2024-2002 2024

  26. [26]

    doi:10.5281/zenodo.5879544 , url =

    Andonian, Alex and Anthony, Quentin and Biderman, Stella and Black, Sid and Gali, Preetham and Gao, Leo and Hallahan, Eric and. doi:10.5281/zenodo.5879544 , url =

    work page doi:10.5281/zenodo.5879544

  27. [27]

    Andrews, Mel , year = 2023, publisher =. The. doi:10.13140/RG.2.2.28311.75685/1 , url =

    work page doi:10.13140/rg.2.2.28311.75685/1 2023

  28. [28]

    Andrews, Mel , year = 2025, month = sep, journal =. The. doi:10.1007/s10670-025-01010-x , url =

    work page doi:10.1007/s10670-025-01010-x 2025

  29. [29]

    Misgendering in

    Ansara, Y Gavriel and Hegarty, Peter , year = 2013, month = aug, journal =. Misgendering in. doi:10.5172/mra.2013.7.2.160 , url =

    work page doi:10.5172/mra.2013.7.2.160 2013

  30. [30]

    and Busby, Ethan C

    Argyle, Lisa P. and Busby, Ethan C. and Fulda, Nancy and Gubler, Joshua and Rytting, Christopher and Wingate, David , year = 2022, month = sep, number =. Out of. arXiv , keywords =:2209.06899 , primaryclass =

    arXiv 2022

  31. [31]

    Why Do Language Models Perform Worse for Morphologically Complex Languages? , booktitle =

    Arnett, Catherine and Bergen, Benjamin , editor =. Why Do Language Models Perform Worse for Morphologically Complex Languages? , booktitle =

  32. [32]

    and Losongco, Anthony and Wasow, Thomas and Ginstrom, Ryan , year = 2000, journal =

    Arnold, Jennifer E. and Losongco, Anthony and Wasow, Thomas and Ginstrom, Ryan , year = 2000, journal =. Heaviness vs. Newness:. doi:10.1353/lan.2000.0045 , url =

    work page doi:10.1353/lan.2000.0045 2000

  33. [33]

    Rethinking Child Difficulty:

    Arnon, Inbal , year = 2010, month = jan, journal =. Rethinking Child Difficulty:. doi:10.1017/S030500090900943X , url =

    work page doi:10.1017/s030500090900943x 2010

  34. [34]

    The Nature of

    Arnon, Inbal , editor =. The Nature of. Trends in. doi:10.1075/tilar.19.07arn , url =

    work page doi:10.1075/tilar.19.07arn

  35. [35]

    Arnon, Inbal , year = 2021, month = sep, journal =. The. doi:10.1017/S0305000921000386 , url =

    work page doi:10.1017/s0305000921000386 2021

  36. [36]

    What Enables Human Language?

    Arnon, Inbal and Carmel, Liran and Claidi. What Enables Human Language?. Science , volume =. doi:10.1126/science.adq8303 , url =

    work page doi:10.1126/science.adq8303

  37. [37]

    Proceedings of the 62nd

    Arora, Aryaman and Jurafsky, Dan and Potts, Christopher , year = 2024, pages =. Proceedings of the 62nd. doi:10.18653/v1/2024.acl-long.785 , url =

    work page doi:10.18653/v1/2024.acl-long.785 2024

  38. [38]

    Understanding

    Arps, David and Sajjad, Hassan and Kallmeyer, Laura , year = 2025, publisher =. Understanding. doi:10.48550/ARXIV.2508.07969 , url =

    work page doi:10.48550/arxiv.2508.07969 2025

  39. [39]

    Uncovering Syllable Constituents in the Self-Attention-Based Speech Representations of Whisper , booktitle =

    A Shams, Erfan and Gessinger, Iona and. Uncovering Syllable Constituents in the Self-Attention-Based Speech Representations of Whisper , booktitle =

  40. [40]

    Askari, Raha and Zarrie. Are. Proceedings of the First. doi:10.18653/v1/2025.babylm-main.4 , url =

    work page doi:10.18653/v1/2025.babylm-main.4 2025

  41. [41]

    Probing Subphonemes in Morphology Models , booktitle =

    Astrach, Gal and Pinter, Yuval , editor =. Probing Subphonemes in Morphology Models , booktitle =

  42. [42]

    Synthese , volume =

    Do Language Models Lack Communicative Intentions? , author =. Synthese , volume =. doi:10.1007/s11229-025-05022-6 , url =

    work page doi:10.1007/s11229-025-05022-6

  43. [43]

    doi:10.35111/GS6S-GM48 , url =

    Baayen, Harald and R, Piepenbrock and L, Gulikers , year = 1995, pages =. doi:10.35111/GS6S-GM48 , url =

    work page doi:10.35111/gs6s-gm48 1995

  44. [44]

    Analyzing

    Baayen, Harald , year = 2008, publisher =. Analyzing

    2008

  45. [45]

    Corpus Linguistics and Linguistic Theory , issn =

    The Wompom , author =. Corpus Linguistics and Linguistic Theory , issn =. doi:10.1515/cllt-2024-0053 , url =

    work page doi:10.1515/cllt-2024-0053 2024

  46. [46]

    Baddeley, Alan and Emslie, Hazel and Nimmo-Smith, Ian , year = 1993, month = feb, journal =. The. doi:10.1111/j.2044-8260.1993.tb01027.x , url =

    work page doi:10.1111/j.2044-8260.1993.tb01027.x 1993

  47. [47]

    Journal of Linguistics , volume =

    Toward a Model of Grammaticality Judgments , author =. Journal of Linguistics , volume =. doi:10.1017/S0022226709990260 , url =

    work page doi:10.1017/s0022226709990260

  48. [48]

    and Cross, Keith , year = 2021, month = jul, number =

    Baria, Alexis T. and Cross, Keith , year = 2021, month = jul, number =. The Brain Is a Computer Is a Brain: Neuroscience's Internal Debate and the Social Significance of the. arXiv , keywords =:2107.14042 , primaryclass =

    arXiv 2021

  49. [49]

    Networks in

    Baronchelli, Andrea and. Networks in. Trends in Cognitive Sciences , volume =. doi:10.1016/j.tics.2013.04.010 , file =

    work page doi:10.1016/j.tics.2013.04.010 2013

  50. [50]

    Distributions in Text , booktitle =

    Baroni, Marco , editor =. Distributions in Text , booktitle =. doi:10.1515/9783110213881.2.803 , url =

    work page doi:10.1515/9783110213881.2.803

  51. [51]

    Baroni, Marco and Bernardini, Silvia and Ferraresi, Adriano and Zanchetta, Eros , year = 2009, journal =. The. doi:10.1007/s10579-009-9081-4 , file =

    work page doi:10.1007/s10579-009-9081-4 2009

  52. [52]

    Baroni, Marco , year = 2022, month = nov, edition =. On the. Algebraic. doi:10.1201/9781003205388-1 , url =

    work page doi:10.1201/9781003205388-1 2022

  53. [53]

    Barski, Conrad , year = 2011, publisher =. Land of

    2011

  54. [54]

    Bates, Elizabeth and MacWhinney, Brian and Caselli, Cristina and Devescovi, Antonella and Natale, Francesco and Venza, Valeria , year = 1984, month = apr, journal =. A. doi:10.2307/1129947 , url =. 1129947 , eprinttype =

    work page doi:10.2307/1129947 1984

  55. [55]

    and Elman, Jeffrey L

    Bates, Elizabeth A. and Elman, Jeffrey L. , editor =. Connectionism and the. Brain. doi:10.1002/9780470753507.ch21 , url =

    work page doi:10.1002/9780470753507.ch21

  56. [56]

    doi:10.1515/9783110203325 , url =

    Computational. doi:10.1515/9783110203325 , url =

    work page doi:10.1515/9783110203325

  57. [57]

    Proceedings of the 2024

    Bauwens, Thomas and Delobelle, Pieter , year = 2024, pages =. Proceedings of the 2024. doi:10.18653/v1/2024.naacl-long.324 , url =

    work page doi:10.18653/v1/2024.naacl-long.324 2024

  58. [58]

    Bavaresco, R

    Bavaresco, Anna and Bernardi, Raffaella and Bertolazzi, Leonardo and Elliott, Desmond and Fern. Proceedings of the 63rd. doi:10.18653/v1/2025.acl-short.20 , url =

    work page doi:10.18653/v1/2025.acl-short.20 2025

  59. [59]

    Our Exagmination Round His Factification for Incamination of Work in Progress , author =

  60. [60]

    Scientific Reports , volume =

    Encoding of Speech in Convolutional Layers and the Brain Stem Based on Language Experience , author =. Scientific Reports , volume =. doi:10.1038/s41598-023-33384-9 , url =

    work page doi:10.1038/s41598-023-33384-9

  61. [61]

    Begu. Large. doi:10.48550/arXiv.2305.00948 , url =. arXiv , keywords =:2305.00948 , primaryclass =

    work page doi:10.48550/arxiv.2305.00948

  62. [62]

    Basic Syntax from Speech:

    Begu. Basic Syntax from Speech:. arXiv , keywords =:2305.01626 , primaryclass =

    Pith/arXiv arXiv

  63. [63]

    Begu. Large. IEEE Transactions on Artificial Intelligence , volume =. doi:10.1109/TAI.2025.3575745 , url =

    work page doi:10.1109/tai.2025.3575745 2025

  64. [64]

    Begu. Latent. Antikythera Digital Journal , doi =

  65. [65]

    Von Deutscher

    Behaghel, Otto , year = 1930, journal =. Von Deutscher

    1930

  66. [66]

    The Acquisition of Argument Structure , booktitle =

    Behrens, Heike , editor =. The Acquisition of Argument Structure , booktitle =. doi:10.1515/9783110198775.2.193 , url =

    work page doi:10.1515/9783110198775.2.193

  67. [67]

    Linguistics , volume =

    Usage-Based and Emergentist Approaches to Language Acquisition , author =. Linguistics , volume =. doi:10.1515/LING.2009.014 , url =

    work page doi:10.1515/ling.2009.014 2009

  68. [68]

    Constructivist

    Behrens, Heike , year = 2021, month = sep, journal =. Constructivist. doi:10.1017/S0305000921000556 , url =

    work page doi:10.1017/s0305000921000556 2021

  69. [69]

    Analyzing

    Beinborn, Lisa and Pinter, Yuval , year = 2023, pages =. Analyzing. Proceedings of the 2023. doi:10.18653/v1/2023.emnlp-main.272 , url =

    work page doi:10.18653/v1/2023.emnlp-main.272 2023

  70. [70]

    Cognitive

    Beinborn, Lisa and Hollenstein, Nora , year = 2024, series =. Cognitive. doi:10.1007/978-3-031-43260-6 , url =

    work page doi:10.1007/978-3-031-43260-6 2024

  71. [71]

    Writing Your Journal Article in Twelve Weeks: A Guide to Academic Publishing Success , shorttitle =

    Belcher, Wendy Laura , year = 2019, series =. Writing Your Journal Article in Twelve Weeks: A Guide to Academic Publishing Success , shorttitle =

    2019

  72. [72]

    Analysis

    Belinkov, Yonatan and Glass, James , year = 2019, month = apr, journal =. Analysis. doi:10.1162/tacl_a_00254 , url =

    work page doi:10.1162/tacl_a_00254 2019

  73. [73]

    Belinkov, Yonatan , year = 2022, month = apr, journal =. Probing. doi:10.1162/coli_a_00422 , url =

    work page internal anchor Pith review doi:10.1162/coli_a_00422 2022

  74. [74]

    and Gleason, Jean Berko , year = 1982, journal =

    Bellinger, David C. and Gleason, Jean Berko , year = 1982, journal =. Sex

    1982

  75. [75]

    Eliciting Latent Predictions from Transformers with the Tuned Lens

    Belrose, Nora and Furman, Zach and Smith, Logan and Halawi, Danny and Ostrovsky, Igor and McKinney, Lev and Biderman, Stella and Steinhardt, Jacob , year = 2023, month = nov, number =. Eliciting. doi:10.48550/arXiv.2303.08112 , url =. arXiv , keywords =:2303.08112 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.08112 2023

  76. [76]

    and Koller, Alexander , title =

    Bender, Emily M. and Koller, Alexander , year = 2020, publisher =. Climbing towards. Proceedings of the 58th. doi:10.18653/v1/2020.acl-main.463 , file =

    work page doi:10.18653/v1/2020.acl-main.463 2020

  77. [77]

    doi: 10.1145/3442188.3445922

    Bender, Emily M. and Gebru, Timnit and. On the. Proceedings of the 2021. doi:10.1145/3442188.3445922 , url =

    work page doi:10.1145/3442188.3445922 2021

  78. [78]

    Bengio, Yoshua and Ducharme, R. A. Journal of Machine Learning Research , volume =

  79. [79]

    Curriculum learning

    Bengio, Yoshua and Louradour, J. Curriculum Learning , booktitle =. doi:10.1145/1553374.1553380 , url =

    work page doi:10.1145/1553374.1553380

  80. [80]

    Proceedings of the National Academy of Sciences , volume =

    At 6--9 Months, Human Infants Know the Meanings of Many Common Nouns , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.1113380109 , url =

    work page doi:10.1073/pnas.1113380109

Showing first 80 references.