acceptodds
Under review as a conference paper at ICLR 2027

Brain-Aligned Intermediate Transformer States Depend on Modality, Stimulus, and Positional Encoding

Abstract

Alignment between Transformers and brain activity has mainly been studied at the layer level, but each layer contains computational states with different properties. Previous work identified intermediate states that align more strongly with brain activity than the commonly used block output, yet the generality of this finding and the mechanisms behind the advantage of particular states remain unclear. Here, we examined brain alignment across within-block states in text, audio, music-generation, and vision Transformers. We quantified this alignment by using these states to predict brain responses to speech, music, and images. We also compared models that share the same basic architecture but use different positional-encoding schemes. We found that intermediate states generally yielded better predictions than the conventional block output, but that the strongest correspondences varied with input modality, stimulus, and positional encoding. Text and audio models favored distinct state families even for the same brain responses, and the preferred states changed between speech and music within the same audio model. In vision, positional encoding reshaped the pattern of state-specific alignment. We further examined the computations and representational content underlying these differences, providing a mechanistic interpretation of why particular states align with auditory and visual cortex. These results suggest that the correspondence between brain responses and states is more diverse than previously assumed, and that this diversity is rooted in the computation and the representational content of each state.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuroscience, brain alignment, fMRI, encoding model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_doHVOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper explores the interpretability of transformer block activations for encoding of fMRI responses with three input stimulus modalities: speech, text, music and vision. The authors quantify the the importance of transformer block states (internal activations) for voxel-wise encoding performance. This approach is different from previous works that predominantly focused on block outputs rather than their activations. The paper concludes that the most predictive internal state differs with respect to the input modality and stimulus type (early vs. higher-order sensory processing).

UPD Nov 14th 2026: score increased to 3.

Critical strengths

The main question of the paper is clearly stated. The novelty element is clearly present as the authors analyse block activations instead of outputs for multiple modalities. The analyses seem appropriate for the problem addressed.

Critical weaknesses
  1. There is no null significance testing for encoding performance. Since the main results only report the state/block ranking within models per parcel, it is hard to get an idea of how good the encoding performance is on its own or to which extent it is linked to the model pretraining conditions.
  2. The MRI data were preprocessed with different pipelines. For the speech vs. music comparison the authors attribute the difference in performance to the properties of model blocks. However, the heterogeneity harbored within the data can alone account for any difference in encoding performance: in LPP the encoding is reported for group average and the preprocessing for LPP and MGND is different. This is a major result validity concern on my end.
  3. The authors claim that block-wise analyses can guide architectural choices such that sensory processing is captured better. This is a fair statement overall. However, the authors provide no architectural recommendations for future works. It is unclear how the field should benefit from these results concretely.
  4. My suspicion is that some states are strongly correlated within models.
Questions for authors
  1. The encoding is done in the voxel space but the results are reported within HCP-MMP parcels. Why this approach? Why not do the encoding on the parcel level to begin with?
  2. The encoding is done within-subject (only individual runs are held out). Do the authors have an estimation of how much is lost in performance when a whole subject is left out for evaluation?
  3. Could the authors add a null significance test where the link between the pretrained states and the target (fMRI response) is severed? For instance, one could rerun the analyses on randomly initialized models. If computationally feasible, it would be interesting to see a null distribution from ~100 seeds (a single inference pass per seed).
  4. What does the blue-red colorbar stand for in winning ratio histograms?
  5. Could the authors provide absolute encoding accuracies? The aim here would be to have an idea how well an individual state recovers fMRI responses.
4 comments
AuthorsOpenReview ↗

Response #1

We thank the reviewer for the careful evaluation of our paper and for the constructive feedback. We have revised the manuscript and added the analyses described below.

Q1

There is no null significance testing for encoding performance. Since the main results only report the state/block ranking within models per parcel, it is hard to get an idea of how good the encoding performance is on its own or to which extent it is linked to the model pretraining conditions.

A1. We agree. We added null significance tests and found that prediction accuracy was statistically significant for all states in all datasets and models (Appendix I.5). As the reviewer notes, our main results concern the ranking of states; in line with this, we slightly adjusted the wording (e.g., in the Abstract) so that statements comparing intermediate states with block output are phrased in terms of this ranking.

Q2

The MRI data were preprocessed with different pipelines. For the speech vs. music comparison the authors attribute the difference in performance to the properties of model blocks. However, the heterogeneity harbored within the data can alone account for any difference in encoding performance: in LPP the encoding is reported for group average and the preprocessing for LPP and MGND is different. This is a major result validity concern on my end.

A2. We agree. We therefore added a control in which MGND responses were averaged across participants before fitting the encoding models, bringing its response aggregation closer to LPP. The results were highly similar to those of the participant-wise analysis, and the main findings were reproduced under this control; for example, post-rotation Q/K did not become preferred in the acoustic RoPE model (Appendix I.6). Nevertheless, the observed speech–music difference may still reflect other dataset factors, such as acquisition and preprocessing, which we now note in the Discussion and the Limitations section.

Q3

The authors claim that block-wise analyses can guide architectural choices such that sensory processing is captured better. This is a fair statement overall. However, the authors provide no architectural recommendations for future works. It is unclear how the field should benefit from these results concretely.

A3. Thank you for this helpful comment. We now realize that we should have made the practical implications more explicit. For neuroscience, our analyses help relate computational stages within transformers to the functional properties of sensory cortex, as illustrated by the correspondence with independently measured pRFs. For applications in brain-response prediction, the results guide the choice of representations for particular cortical regions—for example, using post-rotation Q/K as features for predicting early visual responses.

For future work aimed at developing more human-like sensory representations, our results suggest possible directions for exploring model design. For example, the visual findings motivate examining how positional encoding shapes spatial representations and their correspondence with human visual cortex. We have revised the Discussion to state these implications and directions for future research explicitly.

Q4

My suspicion is that some states are strongly correlated within models.

A4. We agree. When correlated states have similar prediction accuracy, small fluctuations can affect which state ranks first. We therefore added 95% bootstrap confidence intervals to the winning ratios (Figure 2 for LPP, Figure A4 for MGND, and Figure A6 for NSD), together with statistical comparisons of winning-ratio differences across states, datasets, and ROI groups (Sections 4.1–4.3; Appendix I.2). These analyses quantify the variability of state preferences and the robustness of the reported differences. For example, the early–higher visual difference in relative Q/K-over-FFN preference was positive in all eight comparisons, ranging from 21.2 to 27.8 percentage points.

AuthorsOpenReview ↗

Response #2

Q5

The encoding is done in the voxel space but the results are reported within HCP-MMP parcels. Why this approach? Why not do the encoding on the parcel level to begin with?

A5. Thank you for this question. We now realize that we should have explained this approach more clearly. Our aim is to characterize neural representations, so we fit the encoding models at the finest spatial unit available in fMRI, the voxel: voxels within a parcel can differ in response selectivity, and averaging them before fitting would discard these differences. For reporting, we then summarize the voxel-wise results within ROIs, which are established functional units of cortex. Summarizing voxel-wise encoding results within ROIs is a widely used approach (e.g., Nishimoto et al., 2011; Güçlü & van Gerven, 2015; Kell et al., 2018; Lescroart & Gallant, 2019). Cortical anatomy and functional organization vary across individuals, making direct voxel-to-voxel comparisons across participants difficult, and HCP-MMP parcels provide a common regional framework for summarizing and comparing results. We have clarified this rationale and added these references in the Discussion.

Q6

The encoding is done within-subject (only individual runs are held out). Do the authors have an estimation of how much is lost in performance when a whole subject is left out for evaluation?

A6. We agree that generalization across participants is important. Our evaluation focuses on prediction of held-out stimuli, following common practice in voxel-wise encoding studies (e.g., Nishimoto et al., 2011; Güçlü & van Gerven, 2015; Kell et al., 2018; Lescroart & Gallant, 2019). Generalization to an entirely new participant additionally concerns how stimulus–response relationships are shared across individuals and how their neural response spaces correspond. These are research questions in their own right, as illustrated by work on functional alignment (e.g., Haxby et al., 2011 and Chen et al., 2015), and extend beyond the scope of our comparison of intermediate representations. We have now explicitly acknowledged this in the Discussion.

Q7

Could the authors add a null significance test where the link between the pretrained states and the target (fMRI response) is severed? For instance, one could rerun the analyses on randomly initialized models. If computationally feasible, it would be interesting to see a null distribution from ~100 seeds (a single inference pass per seed).

A7. Thank you for this suggestion. We added null randomizations that disrupt the correspondence between fixed held-out predictions and measured responses. We showed that prediction accuracy was statistically significant for all states in all datasets and models (Appendix I.5).

We appreciate the suggestion about initialized models, which would help clarify the contribution of pretraining to brain alignment. Our main question concerns how brain alignment varies across intermediate states and conditions in pretrained models. Investigating how these patterns arise through pretraining is an interesting further question that extends beyond the focus of the present study. We have therefore highlighted the proposed comparison across multiple random initializations as a valuable direction for future work in the revised Discussion.

Q8

What does the blue-red colorbar stand for in winning ratio histograms?

A8. We now realize that the color scheme was not sufficiently explained, and we apologize for the confusion. We adopted the blue-to-red color scheme used in Mind’s Transformer for consistency with this relevant prior work. The colors identify states in their displayed within-block order; they do not encode accuracy or statistical significance. Bar length indicates the winning ratio, and solid versus hatched bars distinguish early and higher ROI groups. We have clarified these conventions in the captions and legends of Figure 2.

Q9

Could the authors provide absolute encoding accuracies? The aim here would be to have an idea how well an individual state recovers fMRI responses.

A9. Thank you for this suggestion. Appendix I.5 now reports absolute held-out Pearson r values for the most frequently selected intermediate state in each dataset, model, and principal ROI group, alongside the null significance tests described in A1. Representative r values in early sensory ROIs were 0.371 for LPP/Llama post-rotation Q (99th percentile of the null distribution: 0.095), 0.389 for MGND/MusicGen Q (0.086), and 0.251 for NSD/DeiT III RoPE-Mixed post-rotation Q (0.012). Previous encoding studies have demonstrated the predictive value of language, audio, and vision model representations on these datasets (Chen & Sivakumar, 2026; Denk et al., 2026; Conwell et al., 2024). Our results confirm that this established encoding approach captures meaningful response variation in the present setting, providing a sound basis for our comparisons of intermediate states.

Reviewer_doHVOpenReview ↗

Response to rebuttal

The authors addressed my concerns to an extent. I will raise my score to 3.

I suspect, in the future the number of submissions on applications to cognitive / neuroscience will increase. With that in mind, I'd like to urge the authors to pay more attention to preprocessing and acquisition disparities across datasets as they develop new methods in the future. Also, clear presentation and interpretation of the results in the context of brain science will increasingly be seen as relevant as more reviewers with a background in neuroscience will partake in review processes for major ML conferences. Gaps on these fronts are enough to spark doubt about paper soundness in the neurosci community. I am looking forward to reading the revised article.

Thank you again for your thoughtful reassessment and constructive feedback, which have helped us improve the manuscript.

We have explicitly acknowledged the limitations arising from acquisition and preprocessing differences across datasets and clarified the scope of our interpretations in the revised manuscript. We will continue to pay careful attention to these issues in our future work.

Reviewer_hbtZOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper gives a good exploration of the alignment between Transformers and Brain signals. The intermediate states from text, audio, music-generation, and vision models are compared with voxel-wise encoding frameworks. Results show varied outcomes across different modalities, while position embeddings also play a role. The internal representations were also analyzed to interpret state-specific brain alignment.

Critical strengths
  1. The paper is well organized and easy to follow.
  2. Extensive experiments support the contributions.
Critical weaknesses
  1. There is only one encoding model involved in the experiments. It would reduce the generalizability of the results of the analysis.
  2. There is no clear statistical analysis to support the comparison results, which only show the diverse distributions of different tasks but no unified evidence.
Questions for authors

Apart from the weaknesses, there are several questions:

  1. A table or figure summarizing the improvement achieved by the best intermediate state over the block output would make the basis motivation clearer.
  2. Are the winning-ratio results sensitive to representation dimension, ridge regularization, or other factors?
  3. Could you give a clear discussion on the comparison across modalities and intermediate?
  4. How different position embeddings affect the encoding performance, and give evidence for neural processing?
4 comments
AuthorsOpenReview ↗

Response #1

We thank the reviewer for the careful evaluation of our paper and for the constructive feedback. We have revised the manuscript and added the analyses described below.

Q1

There is only one encoding model involved in the experiments. It would reduce the generalizability of the results of the analysis.

A1. Thank you for raising this point. We now realize that the original manuscript did not explain the choice of a single encoding model or its implications for generalizability. Linear readouts, particularly ridge regression, are standard and widely used in voxelwise encoding studies (e.g., Dupré la Tour et al., 2025). Using the same readout provides a consistent comparison of representations across states, models, and datasets, and connects directly to our spatial-readout analysis. We now clarify that our conclusions concern information accessible under this readout; generalization to other readout families remains to be tested (Discussion; Appendix H).

Q2

There is no clear statistical analysis to support the comparison results, which only show the diverse distributions of different tasks but no unified evidence.

A2. We agree. We added two statistical analyses that cover all three datasets. First, null significance tests showed that prediction accuracy was statistically significant for all states in all datasets and models (Appendix I.5). Second, we added a common bootstrap analysis of winning-ratio differences and their changes across datasets or ROI groups (Sections 4.1–4.3; Appendix I.2). The bootstrap results provide statistical support for three main patterns: (i) FFN preference over Q/K in early auditory cortex for the acoustic RoPE model, (ii) a shift toward FFN states for music relative to speech, and (iii) a stronger relative preference for Q/K over FFN in early than in higher visual cortex.

PatternWhat was compared (W: winning ratio of a state; ΔW: difference in W between two states, in percentage points)Result [95% bootstrap interval]
(i) wav2vec2-RoPE, early auditory cortex (LPP)ΔW = W(FFN1 Activated) − W(post-rotation Q), and the same for post-rotation K+20.9 [12.6, 28.6] for Q and +18.9 [9.6, 27.2] for K; both intervals exclude zero
(ii) wav2vec2 and HuBERT, early and higher auditory cortexΔW = W(FFN Activated) − W(state favored for speech: Pre-Attn Norm or block output); tested quantity: ΔW in MGND − ΔW in LPPPositive in all 6 contrasts: +11.5 to +18.2 in five, whose intervals exclude zero, and +6.3 [−10.5, 22.5] in the sixth (wav2vec2, early auditory cortex)
(iii) Four RoPE vision models (NSD)ΔW = W(post-rotation Q or K) − W(FFN Activated); tested quantity: ΔW in early − ΔW in higher visual cortexPositive in all 8 contrasts: +21.2 to +27.8; all eight intervals exclude zero

Q3

A table or figure summarizing the improvement achieved by the best intermediate state over the block output would make the basis motivation clearer.

A3. We agree. We added a summary table comparing the winning ratios of block output and the best intermediate state across datasets, models, and ROI groups (Appendix I.2). The selected state had a higher winning ratio than block output in almost all (32 of 34) conditions, as summarized by dataset below.

DatasetConditions (model × ROI group) in which the selected intermediate state had a higher winning ratio than block outputΔW = W(selected intermediate state) − W(block output), in percentage points: median (range)
LPP8 of 1010.7 (−2.7 to 61.7)
MGND10 of 105.6 (3.2 to 10.1)
NSD14 of 1412.3 (7.6 to 29.6)

We confirmed that the same pattern was obtained when the state was selected by cross-validation within the training data.

Q4

Are the winning-ratio results sensitive to representation dimension, ridge regularization, or other factors?

A4. Thank you for this question. The added PCA controls preserve the principal state-preference patterns after reducing differences in feature count (Appendix I.1). The ridge-parameter audit identified upper-boundary accumulation in some conditions, so we extended the upper bound from 10¹² to 10²² in targeted checks using one representative model and participant per dataset (for MGND and NSD, the participant with the highest upper-end selection rate): Llama on LPP, MusicGen on MGND, and DeiT III on NSD. Across the six targeted conditions, the most frequent winning state was unchanged, and no state's winning ratio changed by more than 0.0051 percentage points (Appendix I.4). Bootstrap intervals and state-reselection rates additionally quantify evaluation-sample variability (Appendices I.2–I.3).

AuthorsOpenReview ↗

Response #2

Q5

Could you give a clear discussion on the comparison across modalities and intermediate?

A5. Thank you for this suggestion. We now realize that we should have connected the comparisons more clearly. Previous within-block studies analyzed only text-input language models (see also our response A3 to Reviewer LsT7). Whether their conclusions hold in other modalities is not obvious: each state performs a different computation, and what it computes depends on the input. This is clearest for positional encoding, because positions index tokens in text, time frames of a waveform in audio, and two-dimensional patch locations in images. The role of each state, including Q/K rotated by RoPE, and its correspondence with the brain therefore need not be the same across modalities. Our comparisons test this directly, and the earlier language-model findings transferred only in part:

  • Text vs. audio models for the same speech responses. The language-model post-rotation Q/K preference does not transfer uniformly to acoustic models, even when they use RoPE.
  • Speech vs. music for the same acoustic models. The models shift toward FFN states in MGND relative to LPP.
  • Vision. RoPE models favor post-rotation Q/K in early visual cortex, whereas higher visual cortex generally favors FFN states. The early-visual pattern is supported by the spatial-readout analysis and now replicated, together with the pRF result, in the THINGS-fMRI dataset (Appendix I.7).

The statistical comparisons supporting these patterns are summarized in A2. The revision thus emphasizes two contributions: testing the transfer of earlier language-model findings and explaining a visual relationship that generalizes to an independent dataset. We have revised the Abstract, Introduction, and Discussion accordingly.

Q6

How different position embeddings affect the encoding performance, and give evidence for neural processing?

A6. Thank you for this question. As for differences among positional-encoding schemes, the marked effect in our data was that of using RoPE instead of APE in vision; finer differences between schemes had little effect. In DeiT III, post-rotation Q and K had winning ratios of 31.2% and 29.9% in early visual cortex with RoPE-Mixed, whereas no state exceeded 14.6% with APE. By contrast, the two RoPE variants (Axial and Mixed) showed the same early-visual preference in both DeiT III and Swin. Likewise, wav2vec2-RelPos and wav2vec2-RoPE, which share the same basic architecture and training dataset but use relative positional encoding and RoPE, respectively, showed similar state preferences: FFN1 Activated was the most frequently selected state in early auditory cortex for speech in both models, and their winning-ratio profiles correlated at r = 0.81–0.98 (Figure 3b). Characterizing how such finer differences in positional encoding affect brain alignment is an interesting direction for future work.

Reviewer_hbtZOpenReview ↗

Response to authors

Thanks for your careful response. Most of my concerns have been addressed. I keep my score.

We thank again the reviewer for the thoughtful feedback! We are glad that our response addressed most of the concerns raised.

Reviewer_LsT7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to evaluated whether the alignment between an transformer and brain activity primarily depends on its modality (text, speech), the stimulus content and/or the positional encoding.

For this, the authors used a standard ridge linear model per layer and compare the related brain scores.

The results tend to be quite specific: e.g. the elements that best predict brain actiivity depends on the text, speech and music stimuli; post rotation Q/K are better in some early sensory regions for text/vision models with ROPE but not for audio models.

Critical strengths
  • The paper use a variety of fMRI datasets, which is quite unusual in the field.
  • The paper emphasize an important blindspot in the field "these findings motivate a computationally explicit approach to brain–AI alignment, in which the unit of comparison is not merely a layer but a transformation and the information it makes accessible."
Critical weaknesses

Methods:

  • The dimensionality of the different embedding vary substantially. This means that the ridge regression applied in cases may suffer from different degrees of overfitting and effective capacity.
  • I think there may be an issue with the splitting (but welcome clarifications from the authors): If I understand correctly, the encoding models use held-out stimuli; However the best state is both selected and evaluated on the same test data. Without split-half replication or a separate confirmation set, wouldn't the winning ratios partly reflect selection noise?

Contributions:

  • The main observation was already reported e.g. Kumar et al. (2024), Chen & Sivakumar (2026). The present work extends it across modalities and model variants, but mainly modulate the previous findings.

  • The results do not lead to a clear picture that will presumably generalize to other datasets and modalities. It is thus difficult to draw a conclusion from it.

Questions for authors
  • What general conclusion(s) should we draw? Can we expect them to generalize to new datasets?

  • Could the results be affected by the dimensionalities of the models?

  • Not really a question, but I would invite the authors to revamp their results and figures to have a more consistent presentation (e.g. of impact of Q,K, V etc). Currently, the approach and presentation varies across modalities, and it reads more like a series of many small results, than a paper focusing on a clear objective and/or providing a clear conclusion.

2 comments
AuthorsOpenReview ↗

Response #1

We thank the reviewer for the careful evaluation of our paper and for the constructive feedback. We have revised the manuscript and added the analyses described below.

Q1

The dimensionality of the different embedding vary substantially. This means that the ridge regression applied in cases may suffer from different degrees of overfitting and effective capacity.

A1. We agree. We added PCA controls that reduce expanded states, such as FFN intermediate activations, to the hidden width of each model and layer. Because the other states already have this width, all states within a layer then share the smallest dimension among them: 1,024 for the wav2vec2, HuBERT, DeiT III, and ViT models, 1,536 for MusicGen, and the stage-specific width for Swin. The only exception is Llama, whose K and V have 512 dimensions, fewer than its hidden width of 2,048; a stricter control therefore reduced all Llama states to 512. The principal winning-ratio patterns were preserved across all dataset–model configurations. The largest change in winning ratio among the PCA-projected states was only 1.7 percentage points. Appendix I.1 reports the original dimensions and these controls. Thus, the principal preferences are not explained by feature count alone.

Q2

I think there may be an issue with the splitting (but welcome clarifications from the authors): If I understand correctly, the encoding models use held-out stimuli; However the best state is both selected and evaluated on the same test data. Without split-half replication or a separate confirmation set, wouldn't the winning ratios partly reflect selection noise?

A2. Thank you for raising this point and for inviting clarification. We now realize that the manuscript did not clearly explain how states are selected and evaluated. All encoding models were fitted and their regularization parameters selected using training data only. Winning ratios summarize the out-of-sample performance of these separately fitted models. The test-set winners are not used for any subsequent fitting or tuning.

We nevertheless agree that winning ratios can vary with the evaluation sample. We therefore added bootstrap intervals and reselection rates, that is, how often the same top state is reselected across bootstrap resamples (Appendices I.2–I.3). Selection was stable overall: the median reselection rate across the 34 dataset–model–ROI conditions exceeded 95%, and rates were at least 80% in 13 of 14 NSD conditions, although some auditory conditions were less stable. When the state was instead selected by cross-validation within the training data, without using the test data, the same state was selected in 27 of 34 conditions, and the selected state had a higher winning ratio than block output in 32 of 34 conditions (Appendix I.2).

AuthorsOpenReview ↗

Response #2

Q3

The main observation was already reported e.g. Kumar et al. (2024), Chen & Sivakumar (2026). The present work extends it across modalities and model variants, but mainly modulate the previous findings.

A3. We now realize that the original manuscript did not clearly state our contribution beyond the previous studies. Figure 1 contrasts our study with previous work such as Chen & Sivakumar (2026), but the text did not make explicit what is new or why it is not a straightforward extension.

The previous studies cover a narrow setting. They established substantial within-block functional differentiation, including systematic comparisons and computational interpretations, but only for text-input language models and brain responses to spoken narratives: Kumar et al. (2024) examined attention-head transformations in BERT and GPT-2, and Chen & Sivakumar (2026) examined within-block states in LLMs. Kumar et al. themselves note that their text-based models sidestep the acoustic features of speech and point to models operating on the speech signal as future work.

Whether their conclusions hold beyond this setting is not obvious. These studies start from the premise that the components of a block perform distinct computations. What each of these computations does, however, depends on the input. This is clearest for positional encoding: positions index tokens in text, time frames of a waveform in audio, and two-dimensional patch locations in images. There is therefore no reason to expect the role of each state, including Q/K rotated by RoPE, or its correspondence with the brain to be the same across modalities.

We found that the earlier findings transfer only in part. We extend the previous studies by establishing where their findings transfer beyond language-model settings and where they do not, and we additionally develop a spatial-readout account that yields a specific prediction supported in an independent dataset:

  1. The language-model post-rotation Q/K preference does not transfer uniformly to acoustic models predicting the same speech responses, even with RoPE. Post-rotation Q/K together accounted for a winning ratio of 98.5% in early auditory cortex for Llama, whereas in wav2vec2-RoPE, FFN1 Activated exceeded post-rotation Q and K by 20.9 and 18.9 percentage points. LPP probes characterize the acoustic and linguistic information accessible from audio states.
  2. In vision, by contrast, RoPE models favor post-rotation Q/K in early visual cortex. Our derivation shows how positional rotation enables location-dependent effective readout weights after patch averaging, and independent NSD pRF measurements support the spatial relevance of this account.
  3. A prediction derived from this principle, an early-visual post-rotation Q/K advantage, was also reproduced in an independent dataset, THINGS-fMRI (Appendix I.7).

We have revised the Abstract, Introduction, and Discussion to state explicitly that we extend within-block analyses beyond language models and what this revealed.

Q4

The results do not lead to a clear picture that will presumably generalize to other datasets and modalities. It is thus difficult to draw a conclusion from it.

A4. We agree. As explained in A3, our findings establish the limits of transferring earlier state preferences and identify a specific spatial-readout relationship supported by independent visual replication. The revision connects these conclusions.

Q5

What general conclusion(s) should we draw? Can we expect them to generalize to new datasets?

A5. Thank you for raising this question. Our general conclusions are summarized in A3 and A4. As described in A3, we tested the visual prediction in THINGS-fMRI using all three participants. Both the early-visual Q/K pattern and the pRF-center correspondence replicated (Appendix I.7). In both datasets, post-rotation Q and K were the two most frequently selected of the 13 states of DeiT III RoPE-Mixed in early visual cortex. These results provide independent evidence that these specific findings extend beyond NSD.

Q6

Could the results be affected by the dimensionalities of the models?

A6. Thank you for this question. The PCA controls described in A1 directly address differences in feature count: the principal patterns persist after reducing expanded states to their model’s hidden width (Appendix I.1).

Q7

Not really a question, but I would invite the authors to revamp their results and figures to have a more consistent presentation (e.g. of impact of Q,K, V etc). Currently, the approach and presentation varies across modalities, and it reads more like a series of many small results, than a paper focusing on a clear objective and/or providing a clear conclusion.

A7. We agree. The revised Introduction, Results, and Discussion connect the limits of earlier findings, representational explanations, and independent visual replication into a more consistent story.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal summary

We thank the reviewers for their constructive feedback. The reviews highlighted the importance of examining intermediate computations beyond block outputs, the breadth of our evaluation across modalities and fMRI datasets, and the extensive experiments supporting our contributions. They also requested clearer interpretation, additional controls, and stronger statistical evidence.

In response, we added independent replication in THINGS-fMRI, controls for dimensionality, state selection, regularization, and response aggregation, and expanded statistical analyses. These additions provide stronger support for our main findings. We also clarified our contribution beyond prior language-model studies: identifying where intermediate-state preferences generalize across modalities and where they do not, and explaining the visual Q/K preference through a spatial-readout account supported by independent replication. Together, these revisions strengthen the evidence for our conclusions and clarify their significance for understanding and predicting sensory brain responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.