acceptodds
Under review as a conference paper at ICLR 2027

Preserving What Matters: Human-Relevant Manifold Geometry of Speech Representations Under Acoustic Intervention

Abstract

Robust representations should preserve what matters across perturbations. Yet invariance alone cannot determine whether a representation has isolated task-relevant identity or simply discarded information required downstream. Here, we introduce functional geometry under intervention, a framework for testing when representational geometry is functionally meaningful. We hold sentence identity fixed while manipulating acoustic realization through clean speech, noise, and two speech-enhancement interventions that produce opposite effects on human comprehension. We characterize sentence manifolds across 429 frozen model-layer representations and test whether their geometry preserves sentence identity under an acoustic condition excluded from representation construction. We find that late Whisper encoders develop a coordinated geometric regime with greater inter-sentence separability and reduced within-sentence intervention radius and effective dimension. Across seven direct-audio encoders, these geometric properties reliably track leave-one-condition-out identity retrieval, and the same pattern is reproduced by mean reciprocal rank. Pairwise transfer further reveals an important distinction hidden by aggregate robustness measures: some representations preserve identity only between acoustically neighboring conditions, whereas others support identity across the full intervention space. Human fMRI provides an independent constraint. Beneficial enhancement selectively shifts Inferior Frontal Gyrus (IFG) representations toward the geometry evoked by clean speech, while cross-intervention identity preservation and cortical predictivity remain related but non-equivalent. Together, these findings show that robust representations cannot be identified from invariance alone. A representation is functionally robust when the object survives an unseen intervention and remains available in a form useful to a downstream system.

Primary area:
applications to neuroscience & cognitive science
Keywords:
representation learning, speech representations, manifold geometry, acoustic interventions, out-of-distribution generalization, human cortical representations, fMRI

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_j9MHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces functional geometry under intervention, a framework for deciding when the geometric structure of a representation is functionally meaningful. The authors hold sentence identity fixed while varying its acoustic realization across clean speech, −3 dB noise, and two speech-enhancement variants that have opposite effects on human comprehension. They characterize sentence manifolds with MFTMA across 429 frozen model-layer representations, and validate the geometry against leave-one-condition-out identity retrieval, 4×4 pairwise condition-transfer matrices, and linear model-to-brain encoding in a 25-participant fMRI experiment. The main findings are that late Whisper encoders develop a coordinated untangling regime whose geometry predicts held-out identity preservation, that pairwise transfer separates representations preserving identity only locally from those generalizing across the full intervention space, and that beneficial enhancement selectively shifts the inferior frontal gyrus toward the clean-speech direction.

Critical strengths

S1: The pairwise-transfer analysis exposes a distinction hidden by aggregate invariance measures. Some models (wav2vec2, WavLM) preserve sentence identity only between acoustically neighboring conditions, while others (Whisper) generalize across the full intervention space. This is a genuinely informative characterization of existing models and gives the community a reusable diagnostic for telling apart "isolated identity" from "discarded information."

S2: The design is methodologically careful. IFG was pre-specified as the cortical target before model-to-brain fitting, all condition-level analyses are strictly train-three/test-one, and the beneficial/harmful intervention is defined by recovered meaning rather than perceived quality. The authors are also explicit about which comparisons are descriptive rather than inferential.

Critical weaknesses

W1: The model-level claims are descriptive, not causal. The local-vs.-general transfer gap confounds architecture, scale, and training objective simultaneously, and the geometry-to-function inference rests on only seven model clusters (21 stage summaries) with wide bootstrap intervals. As it stands, the paper re-describes known models with a more informative lens but does not explain why they differ, which is the main reason my score stays at weak accept.

W2: The core geometric quantities are estimated from only four acoustic realizations per sentence, so radius and effective dimension are a coarse proxy for the true nuisance-subspace complexity. Combined with the small number of independent model clusters, the geometry-to-function claim is based on only 21 data points (7 models × 3 stages).

Questions for authors

Q1: If the fMRI section were removed, what remains as a standalone contribution to the ML community? The local-vs.-general distinction appears to stand on its own, but the paper currently presents the cortical constraint as an independent line of evidence.

Q2: Is there any evidence that could attribute the local-vs.-general gap to the weakly-supervised ASR objective rather than scale or architecture, or is the comparison intended to remain strictly descriptive?

Q3: Adding more acoustic realizations per sentence would make the effective dimension closer to true nuisance dimensionality. You already flag this as a limitation, so I note it only as the natural next step.

4 comments
AuthorsOpenReview ↗

Response to Reviewer's Question 1 and Question 2

We sincerely thank the reviewer for the careful evaluation, valuable feedback, and encouraging assessment of our work. We particularly appreciate the recognition of our pairwise-transfer analysis as a reusable diagnostic for distinguishing local from broader identity preservation, as well as the methodological rigor of our experimental design, including behaviorally grounded acoustic interventions, held-out evaluation, and the pre-specified IFG analysis. We will clarify the reviewer's key questions and suggestions below.

Question 1. If the fMRI section were removed, what remains as a standalone contribution to the ML community? The local-vs.-general distinction appears to stand on its own, but the paper currently presents the cortical constraint as an independent line of evidence.

Our work makes three standalone contributions to the ML community, independent of the cortical encoding analysis.

1. Functionally grounded manifold evaluation under acoustic interventions. We extend mean-field theoretic manifold analysis (MFTMA) to sentence-level speech representations under four controlled acoustic conditions. Unlike conventional noise-robustness evaluations based primarily on SNR, our framework includes two practical enhancement methods with experimentally demonstrated opposite effects on comprehension: neural enhancement (SEMamba) improves comprehension, whereas classical enhancement (Ephraim–Malah MMSE) reduces it. This provides a behaviorally grounded setting for examining whether changes in manifold geometry accompany meaningful preservation of sentence identity.

2. Linking manifold geometry to held-out functional generalization. We connect geometric descriptors to leave-one-condition-out (LOCO) sentence retrieval, testing whether representations preserve identity under an unseen acoustic condition. Across seven direct-audio encoders at three processing stages, higher manifold capacity, smaller radius, and lower effective dimension are associated with stronger held-out retrieval (ρ=.643,−.604,−.573\rho = .643, −.604, −.573; FDR-corrected q=.012,.012,.021q = .012, .012, .021; Figure 4). Because these descriptors are correlated, our findings support a coordinated geometric regime rather than identifying a single dominant factor. The key contribution is evaluating geometry through functional generalization, rather than treating geometric invariance as sufficient evidence of robustness.

3. Directional pairwise transfer distinguishes local from broader robustness. Our 1212 directed condition-to-condition evaluations reveal differences hidden by aggregate LOCO performance (Figure 5; Appendix A.2, Figure A3). Some representations preserve sentence identity between acoustically similar conditions but fail across more challenging interventions, whereas others generalize more broadly. This provides a layerwise diagnostic for identifying condition-specific stability versus broader cross-intervention identity preservation.

Together, these contributions establish a standalone ML framework connecting controlled acoustic interventions, representational geometry, and held-out functional performance. Human fMRI adds a complementary assessment of cortical relevance, but is not required to establish these ML findings.

Question 2. Is there any evidence that could attribute the local-vs.-general gap to the weakly-supervised ASR objective rather than scale or architecture, or is the comparison intended to remain strictly descriptive?

Our study characterizes differences in cross-condition generalization across model families but does not attribute them causally to architecture or training objectives. The models differ in architecture, scale, supervision, and training data, preventing these factors from being isolated.

Accordingly, the local-versus-general distinction is an empirical characterization of representational behavior, not a claim about its architectural cause or universality. We will clarify this limitation in Section 5.2. Future controlled comparisons, such as pretrained versus fine-tuned wav2vec2 or WavLM, could help determine which training factors contribute to the observed differences between local and broader generalization.

AuthorsOpenReview ↗

Response to Reviewer's Question 3

Question 3. Adding more acoustic realizations per sentence would make the effective dimension closer to true nuisance dimensionality.

The four-condition scope (K=4K = 4) is explicitly stated in Sections 3.3 and 5.2. Within the practical constraints of human fMRI experiments, we made every effort to design and control a meaningful set of acoustic interventions while maintaining sufficient trials per participant. Our priority was to apply the same controlled interventions to human participants and computational speech models, enabling direct comparisons of their representational geometry under matched acoustic conditions. The four conditions were carefully selected to capture distinct acoustic transformations with contrasting effects on human comprehension, balancing experimental feasibility with scientific relevance.

Future work can extend the MFTMA and LOCO framework to additional SNR levels, noise types, and speech-enhancement methods, as well as broader model architectures and participant samples. Such extensions will allow us to evaluate scalability and generalizability beyond the four conditions examined here, while further characterizing manifold geometry across a richer intervention space.

Thank you for the detailed responses. I believe my score correctly reflects the paper's state, hence I maintain it.

AuthorsOpenReview ↗

Updated Manuscript

Thank you again for your feedback. We would like to inform you that we have uploaded an updated version of our manuscript.

The revised version improves the clarity, organization, and presentation of the Abstract, Introduction, Related Work, Methods, and Discussion. We have also expanded Appendices C.1, C.3.2, and D.7 to provide additional methodological explanations, experimental details, and statistical sensitivity analyses.

Importantly, our core approaches, experimental design, primary analyses, results, figures, and scientific interpretations remain unchanged. These revisions are intended to improve methodological transparency and reproducibility while preserving our original findings and contributions.

Thank you again for your time and consideration.

Reviewer_S7YvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose "functional geometry under intervention", which claims to determine which aspects of representational geometry are functionally relevant for a speech-audio model layer during speech comprehension. This determination happens under two criteria: (1) the geometry is invariant to some controlled perturbation of the acoustic input that preserves "linguistic identity" (meaning?); (2) the embeddings predict the inferior frontal gyrus of participants listening to the same speech inputs.

Critical strengths

The manifold capacity theory applied to speech models is, I believe, novel.

Critical weaknesses

I am in strong support of rejecting this submission, for the following reasons:

  1. The writing is nearly incomprehensible and clearly AI-generated. For instance, the opening sentence: "Robust representations should preserve what matters across perturbations." What perturbations? What is meant by robust? What matters? Another example: "Twenty-five participants heard 96 approximately 10-word sentences across four acoustic conditions (Clean, DNN-SE, Classic-SE, and unprocessed Noisy), with 24 trials per condition assigned using Latin-square counterbalancing across four runs." (the authors never define DNN-SE, Classic-SE, or what is Latin-square counterbalancing, nor give any further details about the participants nor stimuli). The entire article, unfortunately, reads like these examples. It is a shame because the tools used (Chung et al.'s manifold capacity theory) applied to speech processing are genuinely interesting. I would encourage the authors to do a substantial rewrite of the paper, without using AI tools.

  2. I had problems understanding the contribution, largely due to the incoherent writing. It seems, for example, that a key contribution concerned "identity preservation", but this term was never defined. The connection to encoding models is also not clear. First, how did the authors determine that the goal was only to predict IFG, i.e., what was the experiment used to determine IFG as "downstream"? Second, being able to predict IFG and identity preservation under interventions "are complementary rather than interchangeable". This makes the overall method proposed in the paper seem like two unrelated criteria cobbled together. I think this work can benefit from rethinking the criteria, and controlled perturbations for the leave-one-out study. For instance, given that the authors are collecting fMRI data, they can determine what inputs are invariant for a certain brain region (e.g., Feather et al., 2023), and verify that those invariants also hold (or not) for models.

  3. Missing related work on geometry, layer function, and brain alignment:

Abstraction Induces the Brain Alignment of Language and Speech Models. Cheng et al., (ICML 2026);

Model manifold analysis suggests the human visual brain is less like an optimal classifier and more like a feature bank. Conwell and Bonner (Data on the Mind and Brain Findings, 2026).

Questions for authors

None

9 comments
AuthorsOpenReview ↗

Response to Definitions, Experimental Design, and Robustness

We thank the reviewer for the evaluation and for recognizing the novelty of applying manifold capacity theory to speech representations. We appreciate the opportunity to clarify our scientific contributions and methodological foundations. In the following, we address the reviewer's criticisms, clarify potential misunderstandings, and provide supporting evidence regarding the definitions, experimental design, functional identity preservation, rationale for focusing on IFG, scientific novelty, methodological soundness, use of AI tools, and relevant prior work, with references to the corresponding sections of the original manuscript.

1.Definitions, experimental design, and robustness

We respectfully clarify that the terms and details noted in Weakness 1 were explicitly defined in the submitted manuscript:

  • Acoustic perturbations (Noisy, DNN-SE, Classic-SE): Introduced in the Introduction (lines 068–069) and explicitly defined immediately after the sentence quoted by the reviewer in Section 3.1 (lines 135–139) and Appendix C.1 (lines 791–796) by specifying the exact noise condition and enhancement algorithms (SEMamba and Ephraim–Malah MMSE).

  • Sentence identity preservation, robustness, and what matters: Defined in the Introduction (lines 068–071, 081–083, 097–100), Section 3.4 (lines 183–187), and Appendix C.4 (Eqs. C13–C17). It is clear in context that what matters is anchored by our behavioral comprehension dissociation, identity preservation is formalized as 9696-way sentence prototype retrieval, and robustness is explicitly contrasted against mere geometric invariance as maintaining sentence identity across held-out acoustic conditions.

  • Participants, stimuli, and Latin-square counterbalancing: Detailed in Section 3.1 and Appendix C.1 (lines 788–808), covering the 25 participants, 9696 approximately 1010-word spoken sentences (∼3\sim3 s each; 384384 matched waveforms), trial timing, two-alternative forced-choice comprehension task, five-point quality rating, 33T scanner parameters, and SPM12 single-trial GLM estimation. Our Latin-square counterbalancing (Section 3.1, line 134) presents each sentence once per participant while rotating sentence-to-condition assignments across the four runs to balance item and sequence effects (we will cite Scott et al., 2017 and specify the stimulus corpus in Section 3.1 in the revised version).

Reference

Scott, T. L., Gallée, J., & Fedorenko, E. (2017). A new fun and robust version of an fMRI localizer for the frontotemporal language system. Cognitive Neuroscience, 8(3), 167–176. https://doi.org/10.1080/17588928.2016.1201466

AuthorsOpenReview ↗

Response to Central Contribution

2. Central contribution: From geometry to functional preservation

Our central contribution is to link manifold geometry to functional sentence-identity preservation under held-out acoustic interventions, rather than assuming that geometric invariance implies robustness.

First, late Whisper encoders exhibit coordinated geometric untangling: capacity increases from 0.4060.406 to 0.5360.536, while radius decreases from 2.2772.277 to 1.2941.294 and effective dimension from 2.4632.463 to 2.1962.196 (Figure 3). Across seven direct-audio encoders at three stages, these descriptors are associated with held-out retrieval (ρ=.643,−.604,−.573\rho = .643, −.604, −.573; FDR-corrected q=.012,.012,.021q = .012, .012, .021; Figure 4). Mean reciprocal rank supports the same directional pattern. Given the limited number of model clusters and strong descriptor coupling, we interpret these results as a coordinated geometric regime rather than independent causal effects.

Second, directional transfer reveals differences that aggregate retrieval can obscure. wav2vec2-large encoder (layer 1313) transfers well from Clean to DNN-SE (0.5310.531) and from DNN-SE to Clean (0.5420.542), but poorly from Clean to Noisy (0.0210.021) and from Noisy to Clean (0.0620.062). In contrast, Whisper-medium encoder (layer 2424) achieves stronger transfer from Clean to Noisy (0.5940.594) and from Noisy to Clean (0.8230.823; Appendix A.2, Figure A3).

These findings distinguish local acoustic stability from broader identity preservation, demonstrating why representational geometry must be evaluated through function rather than invariance alone.

AuthorsOpenReview ↗

Response to Concerns About IFG Selection and the Complementary Roles of Identity Preservation and Cortical Encoding

3. How did the authors determine that the goal was only to predict IFG, i.e., what was the experiment used to determine IFG as "downstream"?

We did not assume IFG a priori while ignoring other brain regions. As detailed in Section 3.1 (lines 145–149), Section 4.2 (Figure 2), Appendix A (Figure A1), and Appendix B (Figure B1), we analyzed four pre-specified bilateral cortical systems spanning the canonical auditory-speech hierarchy. In Section 4.2 (Figure 2), we showed that bilateral IFG was the only cortical system that exhibited a statistically significant, FDR-corrected state-space shift toward Clean speech under beneficial vs. harmful enhancement (Clean-directed alignment q=.036q=.036; projected displacement q=.025q=.025; Figure 2A–B). Selecting IFG as the downstream comprehension-sensitive target in Section 4.4 was therefore strictly data-driven by the empirical dissociation. Still, we report the result of the other brain regions in our appendix A (figure A1) to be comprehensive.

4. Being able to predict IFG and identity preservation under interventions "are complementary rather than interchangeable". This makes the overall method proposed in the paper seem like two unrelated criteria cobbled together.

As the paper reviewer provided (Conwell & Bonner, 2026) supports, our argument that identity retrieval and IFG encoding are complementary is consistent with the view that biological cortices act less like optimal MFTMA classifiers and more like rich feature banks. Specifically, Geometry alone cannot establish whether sentence identity is preserved, while successful retrieval alone cannot establish relevance to human cortical processing. For example, the Whisper-large decoder achieves high cross-condition retrieval (up to 88.5% pairwise transfer; Figure 5B), yet shows lower IFG predictivity (Figure 6). Combining held-out retrieval with IFG predictivity lets us separate models that achieve invariance by discarding useful structure from those that preserve information relevant to human comprehension. We will integrate Conwell & Bonner (2026) into Sections 2 and 5.1 to ground this theoretical perspective.

Reviewer_dB4yOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines manifold geometry, leave-one-condition-out sentence retrieval, cross-condition transfer, and fMRI encoding to evaluate speech representations under acoustic interventions. Late Whisper representations show coordinated geometric changes associated with better identity preservation. The study also distinguishes local transfer from broader robustness and shows that identity preservation and cortical predictivity are not equivalent.

Critical strengths

The behaviorally beneficial and harmful interventions provide a meaningful functional basis for evaluating robustness beyond acoustic quality. Combining retrieval, pairwise transfer, and cortical encoding reveals differences that a single robustness metric would obscure.

Critical weaknesses

The central geometry–retrieval association relies on only seven model clusters. Bootstrap intervals for capacity and dimension include zero, and within-model associations vary in sign. This limits confidence that the relationship extends beyond the particular model-family composition. The cortical manifold construction is insufficiently specified. Moreover, the small IFG encoding correlations are difficult to assess without clearly reported participant-level uncertainty and inference results, weakening the human-relevance contribution.

Questions for authors

How sensitive are the geometry–retrieval associations to excluding individual models or model families, and what small-sample correction was used for clustered inference? How exactly are cortical manifolds constructed, and how is sentence composition accounted for across acoustic conditions? Can the authors report participant-level confidence intervals and corrected significance tests for the principal IFG encoding results?

3 comments
AuthorsOpenReview ↗

Response to Reviewer's Questions

We sincerely thank the reviewer for the thoughtful evaluation, constructive feedback, and positive assessment of our work. We particularly appreciate the recognition of our behaviorally meaningful acoustic interventions and the complementary insights provided by retrieval, pairwise transfer, and cortical encoding. We address the reviewer's concerns below.

Q1: How sensitive are the geometry–retrieval associations to excluding individual models or model families, and what small-sample correction was used for clustered inference?

Investigating model sensitivity, we found that the relationship remains stable across all leave-one-model-out folds when excluding any single model (ρ(αM)∈[+0.50,+0.78],ρ(RM)∈[−0.72,−0.49],ρ(DM)∈[−0.75,−0.46]\rho(\alpha_M)\in[+0.50,+0.78], \rho(R_M)\in[-0.72,-0.49], \rho(D_M)\in[-0.75,-0.46]; all p<.05p<.05).

To account for the small-sample issue the reviewer kindly pointed out, we supplemented our rank regressions on the N=21N=21 encoder stage summaries with a wild cluster bootstrap using Webb 66-point weights (Webb, 2023) and a conservative reference distribution (t(6)t(6)). Under the bootstrap, associations with Top-1 identity retrieval remain statistically significant for manifold capacity (ρ=+0.645,pWebb=.022,pt(6)=.006\rho=+0.645,p_{\text{Webb}}=.022,p_{t(6)}=.006), radius (ρ=−0.616,pWebb=.016,pt(6)=.003\rho=-0.616,p_{\text{Webb}}=.016,p_{t(6)}=.003), and dimension (ρ=−0.595,pWebb=.039,pt(6)=.009\rho=-0.595,p_{\text{Webb}}=.039, p_{t(6)}=.009).

Q2: How exactly are cortical manifolds constructed, and how is sentence composition accounted for across acoustic conditions?

To clarify the cortical condition-manifold formulation, we will add the following definition to the Methods section in the revised version:

For each participant ss and ROI rr, each condition cc is represented as a condition-level manifold Xs,c,rfMRI∈RVr′×24X_{s,c,r}^{fMRI}\in\mathrm R^{V^\prime_r\times24} formed by the 2424 single-trial GLM β\beta vectors within a shared top-10% active voxel subspace (Vr′V^\prime_r, selected via orthogonal task-vs.-baseline tt-statistics). Applying MFTMA yields condition-specific radius RM(s,r,c)R_M(s,r,c) and dimension DM(s,r,c)D_M(s,r,c). These are log-transformed and zz-scored across all participant ×\times condition observations within each ROI to form the 2D cortical state vector gs,r,c=[z(log⁡RM),z(log⁡DM)]Tg_{s,r,c}=[z(\log R_M),z(\log D_M)]^T, and averaged into the four pre-specified bilateral cortical systems for further analysis.

Regarding the effect of sentence composition, we fitted a logistic mixed-effects model including both Subject and Sentence random intercepts (Correct ~ Condition + (1|Subject) + (1|Sentence)), and found that even after absorbing sentence-level variance, the condition effect (χ2(3)=32.24,p=4.7×10−7\chi^2(3)=32.24,p=4.7\times10^{-7}), the beneficial vs. harmful enhancement contrast (DNN-SE vs. Classic-SE: Δlog-odds=+1.35,z=2.95,p=.0032\Delta\text{log-odds} = +1.35, z = 2.95, p = .0032; Bayesian general mixed linear model posterior contrasts vs. Noisy: DNN-SE +0.477,p<.001+0.477,p<.001; Classic-SE−0.846,p<.001-0.846, p<.001), and the perceived quality linear mixed model (χ2(3)=135.8,p<2.2×10−16\chi^2(3)=135.8,p<2.2\times10^{-16}) remain highly significant.

Additionally, we compared condition-manifold RMR_M and DMD_M across all 77 models on the exact 44 sets of sentences using either Clean waveforms or their assigned acoustic conditions. Under Clean speech, the intrinsic ΔRM\Delta R_M between the DNN-SE and Classic-SE sentence sets is negligible (ΔRM=0.0030\Delta R_M=0.0030), whereas the actual acoustic interventions induce a 2525 times larger geometric divergence (ΔRM=0.0752\Delta R_M=0.0752). In addition, all late encoder layers show that the subset geometry correlates strongly with the 9696-sentence geometry (r=0.948,p=2.1×10−86r=0.948,p=2.1\times10^{-86} for RMR_M; r=0.794,p=1.4×10−38r=0.794,p=1.4\times10^{-38} for DMD_M), supporting that the differences in these metrics were induced by the acoustic conditions.

Q3: Can the authors report participant-level confidence intervals and corrected significance tests for the principal IFG encoding results?

We conducted a 20,00020,000-permutation paired sign-flip test to obtain pp-values and participant-level bootstrapped CIs, where only 1111 model ×\times stream ×\times stage combinations have CIs that exclude zero and 99 achieve p≤.05p\le.05. When BH-FDR corrected across all 2121 direct-audio stages, no single IFG stage survives (q=.195q=.195). However, this observation is consistent with prior fMRI speech-encoding literature (e.g., Caucheteux et al., 2023) using single-trial GLM β\betas and cross-condition generalization, where absolute correlations in higher-order prefrontal association cortex (IFG) are naturally smaller (r≈.02–.04r\approx .02–.04) due to higher trial-to-trial variability and non-auditory cognitive processes.

Webb, M. D. (2023). Reworking wild bootstrap‐based inference for clustered errors. Canadian Journal of Economics/Revue canadienne d'économique, 56(3), 839-858.

Caucheteux, C., Gramfort, A., & King, J. R. (2023). Evidence of a predictive coding hierarchy in the human brain listening to speech. Nature Human Behaviour, 7(3), 430-441.

AuthorsOpenReview ↗

Updated Manuscript

We would like to inform you that we have uploaded an updated version of our manuscript.

The revised version improves the clarity, organization, and presentation of the Abstract, Introduction, Related Work, Methods, and Discussion. We have also expanded Appendices C.1, C.3.2, and D.7 to provide additional methodological explanations, experimental details, and statistical sensitivity analyses.

Importantly, our core approaches, experimental design, primary analyses, results, figures, and scientific interpretations remain unchanged. These revisions are intended to improve methodological transparency and reproducibility while preserving our original findings and contributions.

Thank you again for your time and consideration.

Thank you for this rebuttal, I will keep my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.