Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Exploring Narrative Clustering in Large Language Models: A Layerwise Analysis of BERT

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read BERT's later layers organize representations by narrative content, while authorial style stays unclustered.

desk verdict A clearly described but confounded empirical study: the content-style comparison cannot support the claim as stated, but the question and transparency make it worth referee time. read the letter →

arxiv 2501.08053 v1 pith:EKUR7FBW submitted 2025-01-14 cs.CL cs.AI

Awritrojit Banerjee , Achim Schilling , Patrick Krauss This is my paper · ORCID
classification cs.CLcs.AI
keywords BERTlayerwiseanalysisnarrativeclusteringauthorialstyleGPT-4neuraltransferdimensionalityreductionGDV
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

This paper tries to show that BERT's internal representations are organized primarily by narrative content rather than by authorial style. Using 1000 GPT-4-generated rewrites of ten literary excerpts, the authors extract the [CLS] token embedding at each of BERT's 13 layers and project them with PCA and MDS. They report that clusters corresponding to the original narratives become progressively tighter and better separated in later layers, while clusters corresponding to the target author styles stay close to zero separation. If true, this supports a hierarchical picture in which BERT's later layers encode semantic content and largely discard fine-grained authorial style. The result matters for interpretability and for using LLMs as model systems in cognitive neuroscience.

What carries the argument

The central object is the layerwise [CLS] token embedding of BERT-base-uncased, extracted for all 13 layers over 1000 narratives. The clustering quality is quantified by the Generalized Discrimination Value (GDV), a z-scored, dimensionality-invariant measure comparing mean intra-class and inter-class Euclidean distances; more negative GDV means stronger separation. The dataset is built by GPT-4 'neural style transfer,' in which each of ten base narratives is rewritten in the style of nine other authors, repeated ten times.

What would settle it

Take the ten base narratives, have human writers or a paraphrasing system produce style variants without GPT-4, recompute GDV for author labels across BERT layers; if author-style GDVs stay near zero, the paper's conclusion survives, and if they become strongly negative, the observed content-style gap was an artifact of GPT-4's rewriting.

Watch

Extended reading notes

Core claim

The central claim is that BERT exhibits strong, increasingly compact clustering by narrative content across layers, with minimal clustering by authorial style, on this corpus. The authors interpret this as BERT prioritizing semantic content over stylistic features and exhibiting localized processing of sequential data analogous to hierarchical abstraction. They also contrast this with prior work where obvious text-type changes (fable, sci-fi, children's story) did produce strong clusters, concluding that BERT is sensitive to gross stylistic transformations but not to the subtle variations of individual authorial voice.

Load-bearing premise

The claim depends on GPT-4's rewriting changing only the author's style and leaving each story's content, wording habits, and information the same.

Editorial extensions

If this is right

  • Later layers of BERT can serve as a content-discriminative feature space for narrative texts, meaning content labels are recoverable from [CLS] embeddings without fine-tuning.
  • Authorial style is largely absent from BERT's [CLS] representation on this corpus, so style-sensitive tasks would need features beyond the sentence-level summary.
  • The gap between content clustering and text-type clustering implies BERT separates coarse genre shifts but not fine-grained writer identity, locating coarse style shifts in later layers.
  • Layerwise GDV trends offer a diagnostic for where semantic abstraction emerges in encoder-only transformers.

Reading between the lines

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

  • Because only the [CLS] token was analyzed, authorial style may still be present in per-token or pooled representations; checking alternative pooling would localize style information more precisely.
  • The tenfold repetition of GPT-4 rewrites could inflate content clusters through shared surface wording; a control using independent content paraphrases would separate genuine semantics from lexical overlap.
  • If the content-over-style bias holds across models, architects could build style-invariant classifiers from later BERT layers while targeting style-transfer objectives at earlier layers or additional heads.
  • The same layerwise clustering analysis applied to decoder-only LLMs would show whether the bias is encoder-specific or a general property of pretraining.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 constructs a corpus of 1000 short narratives by taking opening lines from ten literary works and using GPT-4 to rewrite each excerpt in the style of the other nine authors, with each style transfer repeated ten times. It then extracts [CLS] token embeddings from all 13 layers of BERT-base-uncased, projects them with PCA and MDS, and quantifies cluster separation with the GDV metric. The central claim is that BERT's later layers form strong, progressively compact clusters according to narrative content while showing minimal clustering according to authorial style, indicating that BERT prioritizes semantic content over stylistic features.

Significance. If the central claim were established, the paper would be a useful contribution to layerwise interpretability of encoder-only transformers, complementing prior work on hierarchical abstraction in neural language models. The manuscript has some strengths: it analyzes all 13 layers, uses two complementary dimensionality-reduction methods, and applies a quantitative cluster-separation measure that is parameter-free after defining the class labels. However, the significance is strongly conditional on the validity of the experimental design, and the current evidence does not support the stated conclusion because the content and style manipulations are not cleanly separated and no statistical validation is provided.

major comments (4)
  1. [§II.A.1] The content labels are each defined by exactly one source text written by exactly one author. Consequently, 'narrative content' is perfectly confounded with source identity, author identity, genre, period, and lexical surface form; the observation that content-labelled clusters become compact in later layers cannot be attributed to narrative content per se. The design needs at least two texts per content theme and multiple authors per content condition, or a within-corpus control that varies content while holding the author constant, before the prioritization claim can be tested.
  2. [§II.A.2] The only manipulation of style is an unvalidated GPT-4 prompt asking the model to rewrite each text in another author's style. No human or automated evaluation is reported to show that the rewrites preserve the original narrative content while genuinely exhibiting the target author's style. If GPT-4 also changes informativeness, vocabulary, or framing, the style labels are noisy or contaminated with content differences; repeating the same transfer ten times may further create near-duplicate examples that inflate cluster compactness. The manuscript should report duplicate rates, content-preservation checks, and style-classification accuracy on the generated texts.
  3. [§III, Fig. 5] The paper reports no error bars, confidence intervals, significance tests, or defined baselines for the GDV values. The phrase 'baseline GDVs' is undefined, and the layerwise trend is asserted from a single output of the pipeline. A permutation test with shuffled labels and bootstrapped GDV estimates is needed to establish that the content-versus-style asymmetry and the layerwise decrease are not due to chance, to the specific GPT-4 generations, or to the particular choice of source texts.
  4. [§III, Figs. 1–4] The visual comparison between author-style clusters and content clusters is not quantitative support on its own. Figures 1–4 lack axis labels, color legends, and any display of class overlap or density, so the reader cannot assess how much of the apparent difference is a projection artifact. The GDV values in Fig. 5 need to be accompanied by per-layer distributions or confidence intervals before the claim of a 'progressive' trend is justified.
minor comments (6)
  1. [§II.D] The sentence 'also invariant with respect to a permutation of the components in the N-dimensional data vectors' should refer to D-dimensional vectors, since the data dimension is denoted D.
  2. [§II.C.2] Calling MDS 'parameter-free' is imprecise: classical MDS has no tuning hyperparameters, but the choice of distance metric, the number of output dimensions, and the stress criterion are methodological choices that should be stated explicitly.
  3. [§V] The data and code availability statement says materials 'will be made available upon reasonable request'; for a paper whose main contribution is an empirical measurement, a public repository with the exact GPT-4 prompts, generated texts, and analysis scripts would substantially improve reproducibility.
  4. [§III] The opening sentence 'From Fig.1 and Fig.2, it is clear that no significant clusters emerge' uses the word 'significant' without a statistical test; please rephrase to describe the visual impression and defer inferential language to the quantitative analysis.
  5. [§IV] The relation to prior study [35] should be made explicit in the introduction or methodology: the present paper appears to extend the paradigm of [35] to individual authorial style, and the novelty over that work should be stated.
  6. [II.A.2] The number of unique texts is 100 even after the style transfer is repeated ten times; the paper should state this explicitly and discuss whether the ten repetitions are treated as independent samples or as repeated measures in the analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the content-versus-style clustering result is an empirical measurement using a parameter-free metric, not a derivation that reduces to its inputs.

full rationale

The central result is an observational measurement on BERT's frozen, pre-trained [CLS] embeddings: layerwise GDV values are computed for content labels and style labels from the same embeddings, and the paper reports that content labels yield more negative GDV. The GDV metric [26] is defined by z-scored Euclidean distances and has no free parameters fitted to the present data; citing the authors' earlier metric paper is methodological, not load-bearing. No equation in the paper defines narrative content in terms of the GDV or vice versa, and no parameter is fitted to a subset and then reported as a prediction of a closely related quantity. The unvalidated GPT-4 'neural style transfer' and the asymmetric construction of content versus style classes are legitimate threats to external validity (e.g., weak style transfer or duplicate/repeated narratives could inflate content-cluster compactness), but this is an experimental confound, not circularity: the reported numbers would still be a measurement of the constructed dataset, not a claim that is true by definition. The prior-work citation [35] about clustering by text type is contextual and does not carry the derivation. Thus no circular step can be exhibited; score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No free parameters are fitted to data; the hand-chosen inputs are the number of GPT-4 repetitions, the prompt design, and the author/text selection. The load-bearing assumptions concern GPT-4 style-transfer fidelity and the adequacy of [CLS] embeddings. No new entities are introduced.

free parameters (3)
  • Number of GPT-4 style-transfer repetitions = 10
    Chosen so that 1000 narratives exceed the 768-dimensional embedding size, enabling PCA in column space (Section II.A.2). No principled justification.
  • GPT-4 style-transfer prompt = Unspecified
    The role prompt is described but exact wording, temperature, and model version are not given; the resulting style-transfer quality depends on these choices.
  • Selection of 10 authors and texts = 10 authors, 1 text each
    Hand-picked to span literary modes (Section II.A.1); content classes are as distinct as the authors' works, which directly inflates content clustering.
assumptions (4)
  • domain assumption GPT-4 neural style transfer preserves narrative content while changing only authorial style
    Invoked in Section II.A.2 to define content and style labels; if false, the content/style comparison is invalid.
  • domain assumption The [CLS] token embedding at each layer adequately represents the narrative
    Section II.B.2 extracts only the [CLS] embedding; this is a standard but unverified compression.
  • standard math GDV is a valid measure of cluster separation
    Equation (4) in Section II.D, adopted from the authors' prior work [26]; its validity is taken as established.
  • standard math PCA and MDS projections preserve the cluster structure relevant for visual inspection
    Section II.C relies on standard references [18]-[24] for these techniques.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Narrative Clustering in Large Language Models: A Layerwise Analysis of BERT." pith.science (2026). https://pith.science/paper/EKUR7FBW

@misc{pith2026250108053,
  author       = {Pith},
  title        = {Pith review of: Exploring Narrative Clustering in Large Language Models: A Layerwise Analysis of BERT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EKUR7FBW}},
  note         = {Machine review of arXiv:2501.08053}
}
read the original abstract

This study investigates the internal mechanisms of BERT, a transformer-based large language model, with a focus on its ability to cluster narrative content and authorial style across its layers. Using a dataset of narratives developed via GPT-4, featuring diverse semantic content and stylistic variations, we analyze BERT's layerwise activations to uncover patterns of localized neural processing. Through dimensionality reduction techniques such as Principal Component Analysis (PCA) and Multidimensional Scaling (MDS), we reveal that BERT exhibits strong clustering based on narrative content in its later layers, with progressively compact and distinct clusters. While strong stylistic clustering might occur when narratives are rephrased into different text types (e.g., fables, sci-fi, kids' stories), minimal clustering is observed for authorial style specific to individual writers. These findings highlight BERT's prioritization of semantic content over stylistic features, offering insights into its representational capabilities and processing hierarchy. This study contributes to understanding how transformer models like BERT encode linguistic information, paving the way for future interdisciplinary research in artificial intelligence and cognitive neuroscience.

Figures

Figures reproduced from arXiv: 2501.08053 by the authors.

Figure 1
Figure 1. Layerwise Neural Activations in BERT Projected using MDS when [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 4
Figure 4. Layerwise Neural Activations in BERT Projected using PCA when [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Layerwise Trends in GDV Illustrating Degree of Clustering (lower is [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Probing Internal Representations of Multi-Word Verbs in Large Language Models

    cs.CL 2025-02 reject novelty 4.0 of 10

    A study claims BERT stores phrasal and prepositional verbs in a non-linearly separable way, but its own linear classifiers separate the two categories almost perfectly.

Reference graph

Works this paper leans on

35 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [35]

    Analyzing narrative processing in large language models (llms): Using gpt4 to test bert,

    P. Krauss, J. H ¨osch, C. Metzner, A. Maier, P. Uhrig, and A. Schilling, “Analyzing narrative processing in large language models (llms): Using gpt4 to test bert,” arXiv preprint arXiv:2405.02024 , 2024

    arXiv 2024

  2. [1]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015

    2015

  3. [2]

    Deep learning: A critical appraisal,

    G. Marcus, “Deep learning: A critical appraisal,” arXiv preprint arXiv:1801.00631, 2018

    arXiv 2018

  4. [3]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar et al., “Attention is all you need,” in Advances in Neural Information Processing Systems , 2017, pp. 5998– 6008

    work page 2017

  5. [4]

    Cognitive computational neuro- science,

    N. Kriegeskorte and P. K. Douglas, “Cognitive computational neuro- science,” Nature Neuroscience, vol. 21, no. 9, pp. 1148–1160, 2018

    work page 2018

  6. [5]

    Predictive coding and stochastic resonance as fundamental principles of auditory phantom perception,

    A. Schilling, W. Sedley, R. Gerum, C. Metzner, K. Tziridis, A. Maier, H. Schulze, F.-G. Zeng, K. J. Friston, and P. Krauss, “Predictive coding and stochastic resonance as fundamental principles of auditory phantom perception,” Brain, vol. 146, no. 12, pp. 4809–4825, 2023

    work page 2023

  7. [6]

    Neural network based successor representations to form cognitive maps of space and language,

    P. Stoewer, C. Schlieker, A. Schilling, C. Metzner, A. Maier, and P. Krauss, “Neural network based successor representations to form cognitive maps of space and language,” Scientific Reports, vol. 12, no. 1, p. 11233, 2022

    work page 2022

  8. [7]

    Neural network based formation of cognitive maps of semantic spaces and the putative emergence of abstract concepts,

    P. Stoewer, A. Schilling, A. Maier, and P. Krauss, “Neural network based formation of cognitive maps of semantic spaces and the putative emergence of abstract concepts,” Scientific Reports , vol. 13, no. 1, p. 3644, 2023

    work page 2023

Show all 35 references
  1. [8]

    Multi-modal cognitive maps based on neural networks trained on successor representations,

    ——, “Multi-modal cognitive maps based on neural networks trained on successor representations,” arXiv preprint, 2023

    2023

  2. [9]

    Word class representations spontaneously emerge in a deep neural network trained on next word prediction,

    K. Surendra, A. Schilling, P. Stoewer, A. Maier, and P. Krauss, “Word class representations spontaneously emerge in a deep neural network trained on next word prediction,” arXiv preprint, 2023

    2023

  3. [10]

    Biological constraints on neural network models of cognitive function,

    F. Pulvermuller, R. Tomasello, M. R. Henningsen-Schomers, and T. Wennekers, “Biological constraints on neural network models of cognitive function,” Nature Reviews Neuroscience , vol. 22, no. 8, pp. 488–502, 2021

    2021

  4. [11]

    Sparsity through evolutionary pruning prevents neuronal networks from overfit- ting,

    R. C. Gerum, A. Erpenbeck, P. Krauss, and A. Schilling, “Sparsity through evolutionary pruning prevents neuronal networks from overfit- ting,” Neural Networks, vol. 128, pp. 305–312, 2020

    2020

  5. [12]

    Integration of leaky-integrate-and-fire neurons in standard machine learning architectures to generate hybrid networks: A surrogate gradient approach,

    R. C. Gerum and A. Schilling, “Integration of leaky-integrate-and-fire neurons in standard machine learning architectures to generate hybrid networks: A surrogate gradient approach,” Neural Computation, vol. 33, no. 10, pp. 2827–2852, 2021

    2021

  6. [13]

    Leaky-integrate- and-fire neuron-like long-short-term-memory units as model system in computational biology,

    R. Gerum, A. Erpenbeck, P. Krauss, and A. Schilling, “Leaky-integrate- and-fire neuron-like long-short-term-memory units as model system in computational biology,” in 2023 International Joint Conference on Neural Networks (IJCNN) . IEEE, 2023, pp. 1–9

    2023

  7. [14]

    Coincidence detection and integration behavior in spiking neural networks,

    A. Stoll, A. Maier, P. Krauss, R. Gerum, and A. Schilling, “Coincidence detection and integration behavior in spiking neural networks,”Cognitive Neurodynamics, pp. 1–13, 2023

    2023

  8. [15]

    Can we open the black box of ai?

    D. Castelvecchi, “Can we open the black box of ai?” Nature News, vol. 538, no. 7623, p. 20, 2016

    2016

  9. [16]

    BERT: Pre- training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre- training of deep bidirectional transformers for language understanding,” in Proceedings of NAACL-HLT, 2019, pp. 4171–4186

    2019

  10. [17]

    Explainable artificial in- telligence: Understanding, visualizing and interpreting deep learning models,

    W. Samek, T. Wiegand, and K.-R. M ¨uller, “Explainable artificial in- telligence: Understanding, visualizing and interpreting deep learning models,” arXiv preprint arXiv:1708.08296 , 2017

    2017 arXiv

  11. [18]

    Multidimensional scaling: I. theory and method,

    W. S. Torgerson, “Multidimensional scaling: I. theory and method,” Psychometrika, vol. 17, no. 4, pp. 401–419, 1952

    1952

  12. [19]

    Nonmetric multidimensional scaling: a numerical method,

    J. B. Kruskal, “Nonmetric multidimensional scaling: a numerical method,” Psychometrika, vol. 29, no. 2, pp. 115–129, 1964

    1964

  13. [20]

    J. B. Kruskal and M. Wish, Multidimensional scaling . Sage, 1978, vol. 11

    1978

  14. [21]

    Multidimensional scaling,

    M. A. Cox and T. F. Cox, “Multidimensional scaling,” in Handbook of data visualization. Springer, 2008, pp. 315–347

    2008

  15. [22]

    Sleep as a random walk: a super-statistical analysis of eeg data across sleep stages,

    C. Metzner, A. Schilling, M. Traxdorf, H. Schulze, and P. Krauss, “Sleep as a random walk: a super-statistical analysis of eeg data across sleep stages,” Communications Biology, vol. 4, no. 1, p. 1385, 2021

    2021

  16. [23]

    Extracting continuous sleep depth from eeg data without machine learning,

    C. Metzner, A. Schilling, M. Traxdorf, H. Schulze, K. Tziridis, and P. Krauss, “Extracting continuous sleep depth from eeg data without machine learning,” Neurobiology of Sleep and Circadian Rhythms , vol. 14, p. 100097, 2023

    2023

  17. [24]

    Classification at the accuracy limit: facing the problem of data ambiguity,

    C. Metzner, A. Schilling, M. Traxdorf, K. Tziridis, A. Maier, H. Schulze, and P. Krauss, “Classification at the accuracy limit: facing the problem of data ambiguity,” Scientific Reports, vol. 12, no. 1, p. 22121, 2022

    2022

  18. [25]

    Analysis of continuous neuronal activity evoked by natural speech with computational corpus linguistics methods,

    A. Schilling, R. Tomasello, M. R. Henningsen-Schomers, A. Zankl, K. Surendra, M. Haller, V . Karl, P. Uhrig, A. Maier, and P. Krauss, “Analysis of continuous neuronal activity evoked by natural speech with computational corpus linguistics methods,” Language, Cognition and Neur...

    2021

  19. [26]

    Quantify- ing the separability of data classes in neural networks,

    A. Schilling, A. Maier, R. Gerum, C. Metzner, and P. Krauss, “Quantify- ing the separability of data classes in neural networks,”Neural Networks, vol. 139, pp. 278–293, 2021

    2021

  20. [27]

    Analysis and visualization of sleep stages based on deep neural networks,

    P. Krauss, C. Metzner, N. Joshi, H. Schulze, M. Traxdorf, A. Maier, and A. Schilling, “Analysis and visualization of sleep stages based on deep neural networks,” Neurobiology of sleep and circadian rhythms, vol. 10, p. 100064, 2021

    2021

  21. [28]

    Analysis of structure and dynamics in three-neuron motifs,

    P. Krauss, A. Zankl, A. Schilling, H. Schulze, and C. Metzner, “Analysis of structure and dynamics in three-neuron motifs,” Frontiers in Compu- tational Neuroscience, vol. 13, p. 5, 2019

    2019

  22. [29]

    Recurrence reso- nance

    P. Krauss, K. Prebeck, A. Schilling, and C. Metzner, “Recurrence reso- nance” in three-neuron motifs,” Frontiers in computational neuroscience, p. 64, 2019

    2019

  23. [30]

    Weight statistics controls dynamics in recurrent neural networks,

    P. Krauss, M. Schuster, V . Dietrich, A. Schilling, H. Schulze, and C. Metzner, “Weight statistics controls dynamics in recurrent neural networks,” PloS one, vol. 14, no. 4, p. e0214541, 2019

    2019

  24. [31]

    Quantifying and maximizing the information flux in recurrent neural networks,

    C. Metzner, M. E. Yamakou, D. V oelkl, A. Schilling, and P. Krauss, “Quantifying and maximizing the information flux in recurrent neural networks,” arXiv preprint arXiv:2301.12892 , 2023

    2023 arXiv

  25. [32]

    A statistical method for analyzing and comparing spatiotemporal cortical activation patterns,

    P. Krauss, C. Metzner, A. Schilling, K. Tziridis, M. Traxdorf, A. Woll- brink, S. Rampp, C. Pantev, and H. Schulze, “A statistical method for analyzing and comparing spatiotemporal cortical activation patterns,” Scientific reports, vol. 8, no. 1, pp. 1–9, 2018

    2018

  26. [33]

    Analysis of multichannel eeg patterns during human sleep: a novel approach,

    P. Krauss, A. Schilling, J. Bauer, K. Tziridis, C. Metzner, H. Schulze, and M. Traxdorf, “Analysis of multichannel eeg patterns during human sleep: a novel approach,” Frontiers in human neuroscience , vol. 12, p. 121, 2018

    2018

  27. [34]

    Microstructure of cortical activity during sleep reflects respiratory events and state of daytime vigilance,

    M. Traxdorf, P. Krauss, A. Schilling, H. Schulze, and K. Tziridis, “Microstructure of cortical activity during sleep reflects respiratory events and state of daytime vigilance,” Somnologie, vol. 23, no. 2, pp. 72–79, 2019

    2019

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.