acceptodds
Under review as a conference paper at ICLR 2027

Encoding Without Alignment: Analyzing The Correspondence Between LLMs And The Visual Cortex

10 views

Abstract

Large language model (LLM) embeddings learned only from text can predict human visual-cortical responses to natural scenes, but whether this predictive correspondence implies shared representational geometry remains unclear. Using fMRI responses from all eight Natural Scenes Dataset subjects and Llama-3-8B representations of the corresponding COCO captions, we compare two notions of brain-model correspondence. First, voxel-wise ridge encoding yields positive held-out prediction accuracy across cortical regions, with higher accuracy in body-selective cortex than in early visual cortex and little variation across Llama layers. Second, we characterize local representational geometry within semantic neighborhoods. Principal-component variation patterns over the shared neighboring stimuli define subspaces in a common neighbor-index space, whose orientations we compare using chordal distance. Within each system, tangent-space distance increases with separation along the cortical ROI ordering and with Llama layer separation. Across systems, however, alignment remains close to an isotropic random-subspace baseline, while being modestly stronger in EBA and FBA than in early visual cortex. Local variance spectra also dissociate: cortical effective dimensionality varies comparatively little across regions, while Llama effective dimensionality increases with layer depth. An exploratory residual-stream intervention using encoding-derived directions produces mixed and nonspecific effects across four subjects. In contrast to Llama, CLIP text-encoder representations show increasing cortical encoding accuracy and stronger cross-system alignment with depth. These results demonstrate that accurate linear encoding can coexist with weak alignment of dominant local variation subspaces.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain–LLM Alignment, Representational Geometry, Neural Encoding Models, Visual Cortex, Mechanistic Interpretability, Representation Steering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_HZkmOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper analyzes the relationship between LLM hidden layer activations and the brain fMRI stimulus. The authors first uses regression to test whether linear combinations of LLM activations can predict brain stimulus, and then evaluated subspace orientation similarity between LLM and brain representations. Experiments show that while there exists some linear relation between LLM hidden layer activations and brain stimulus, the principal subspaces are not aligned.

Critical strengths

1.The paper presents a new perspective for the alignment of the LLM and brain representation. 2.Preliminary results support that some linear relation exists between LLM activation and brain stimulus patterns.

Critical weaknesses

1.Significance of conclusion we can derive from the experiments. The experiments show that the LLM representation can be a weak linear predictor of the brain stimulus, but the two subspaces are hardly aligned. However, a general linear mapping need not preserve leading principal-component subspaces. The paper does not sufficiently explain what additional scientific insight the observed dissociation provides beyond this mathematical fact, or how it advances our understanding of the relationship between LLM and cortical representations. 2.Method validity. The study tries to draw correlation between the brain fMRI stimulus from visual input and LLM activation representation from textual input. Since these two modalities hold different information, weak alignment may partly reflect differences in the information available to the two systems, rather than differences in representational organization. 3.Insufficient experiment results. The study includes only one LLM, one text encoder, and one fMRI dataset. Whether the finding generalizes to other models and datasets is not tested. 4.Writing. The authors could enhance the writing by clearly stating the problem, the hypothesis, main findings and conclusions. The current introduction and discussion part is not well-organized and hard to read.

Questions for authors

1.What scientific insight does the observed dissociation provide beyond the fact that linear mappings need not preserve dominant principal-component subspaces? How does this finding refine existing claims about brain–LLM correspondence? 2.Could the authors provide methods to distinguish representational differences from differences between image and caption information? 3.Please show results on more models and fMRI datasets to demonstrate that the claimed relationship consistently exists across models and datasets.

3 comments

We appreciate our reviewer’s recognition that the paper presents a “new perspective for the alignment of the LLM and brain representation.” We address the weaknesses and questions in order below.

W3.1/Q3.1 We agree with our reviewer’s premise that a general linear map need not preserve principal-component subspaces. The question we address is empirical: given the mathematical freedom of linear encoding to use any direction in LLM feature space, how much do their dominant subspaces overlap? Our contribution with this paper is to evaluate the extent of this dissociation in brain-model alignment and propose geometric alignment as a complementary measure of representational correspondence. Previous work has demonstrated correspondence between language-model representations and cortical responses using linear encoding models, including Caucheteux and King (2022) for natural language comprehension and Doerig et al. (2025) for visually evoked cortical responses predicted from image-caption embeddings. While successful prediction indicates that model features are linearly informative about cortical activity, it does not establish that the two systems share similar patterns of variation across stimuli. Our approach addresses this distinction by comparing local tangent spaces in a shared stimulus-coordinate space (defined by the leading left singular vectors), so their alignment is invariant to orthogonal rotations of the original feature coordinates. This allows us to compare the dominant patterns of stimulus variation without requiring direct correspondence between individual model features and cortical voxels. Our framework also allows us to examine how local tangent-space orientations vary across model layers and along the cortical hierarchy, and we show systematic changes in subspace orientation with increasing separation in both systems.

We agree with the reviewer’s premise that a general linear map need not preserve principal-component subspaces. The question we address is empirical: given the mathematical freedom of linear encoding to use any direction in LLM feature space, do the brain and LLM in fact share dominant subspaces when the field’s standard ridge-encoding pipeline is applied? The literature has largely assumed yes, interpreting encoding scores geometrically (e.g., as evidence of ‘shared representations’). Conwell et al. (2024) have already shown this interpretation is problematic: 126 vision models reach OTC encoding scores within r=0.1r = 0.1 despite radical architectural differences. Our tangent-space diagnostic is a specifically stronger criterion that does distinguish brain-model pairs with and without shared principal subspaces. The observation that EBA (the strongest effect) as well as the other regions V1-hV4 sit below the random-subspace baseline is a non-trivial empirical claim: This graded dependence is the new scientific insight: the dissociation is not a universal fact of linear mappings but an ROI-specific finding about which cortical regions share geometry with caption-LLM representations.

W3.2/Q3.2 We thank our reviewer for highlighting this important methodological concern: differences in input modality may contribute to the observed weak brain-model alignment, as the brain processes visual scenes while the LLM receives textual descriptions. Our cross-modal comparison of brain and LLM representations is motivated by prior work, notably Doerig et al. (2025), which demonstrated that language-model representations of scene captions can predict cortical responses to the corresponding viewed images. However, we agree that even when images and their corresponding textual descriptions capture similar semantic content, weak geometric alignment may reflect differences in the information available to each system rather than differences in how they organize shared stimulus information.

The CLIP-text encoder comparison in Appendix S5 directly addresses this concern. Both Llama-3-8B and CLIP-text receive identical caption inputs for the same stimuli, so any difference between their cortical correspondence patterns cannot be attributed to the caption-vs-image modality asymmetry. We observe that CLIP-text shows depth-dependent alignment (monotone encoding, stronger cross-system chordal at deeper layers) while Llama shows layer-flat alignment. Because the two models process identical text inputs, the difference isolates representational-organization effects from modality content. To further strengthen this disentanglement, we have added a Qwen2.5-7B replication on S7. Qwen shows a similar layer-flat encoding pattern as Llama (although its layer dependence is more pronounced ) and a similar cross-system chordal signature, confirming the finding is a property of text-only autoregressive LLMs rather than Llama-specific.

W3.3/Q3.3: We thank the reviewer for this suggestion. We have extended our analysis to Qwen2.5-7B and two additional fMRI datasets, BOLD5000 (four participants) and THINGS-fMRI (three participants), with the results included in Appendices S5--S7 of the revised manuscript.

On NSD, Qwen reproduces the regional alignment pattern observed with Llama: EBA shows the strongest correspondence across all four subjects, followed by FBA and early visual cortex (mean chordal distances of 7.054, 7.174, and 7.269, respectively; random baseline 7.354). Unlike Llama, however, Qwen exhibits depth-dependent alignment, with EBA chordal distance decreasing from approximately 7.25 at layer 4 to 6.95 at layer 24.

The results also extend to BOLD5000 and THINGS-fMRI. On BOLD5000, Llama achieves positive held-out encoding accuracy (r≈0.11r \approx 0.11–0.250.25), with LOC outperforming EarlyVis in all four participants. On THINGS-fMRI, encoding remains positive (r≈0.025r \approx 0.025–0.0890.089), and EBA shows stronger geometric alignment than the average early visual region across all three participants. In both datasets, brain-Llama chordal distances remain below the random-subspace baseline, although the regional ordering is less consistent than in NSD.

One relevant difference is that BOLD5000 and THINGS-fMRI were acquired at 3T, whereas NSD was acquired at 7T. Differences in acquisition resolution and signal reliability may contribute to the weaker correspondence observed in these datasets. However, stimulus sets, caption sources, and ROI definitions also differ, so we cannot isolate the contribution of scanner field strength.

W3.4: We thank our reviewer for highlighting areas where our presentation could be improved. We have revised the Introduction to clarify the central question of whether successful cortical prediction implies shared local representational geometry, motivate tangent-space alignment as a complement to linear encoding, and summarize our principal findings. We have also revised the Discussion to specify the dissociation between cortical encoding and geometric alignment, characterize within-system organization and cross-system correspondence, and evaluate robustness and generalizability using additional Appendix analyses.

Thank you for the response and the additional experiments. The newly added results address my concern on the experimental coverage. The revised presentation also makes the central findings clearer. However, two concerns remain.

  1. The additional scientific insight relative to prior work remains insufficiently established. Doerig et al. (2025) supported brain–model correspondence using linear encoding and other strategies. The observed disassociation between linear predictability and local principal-subspace alignment does not clarify how the present findings refine that work.

  2. The contribution of input-modality differences remains unresolved. Comparing Llama and CLIP-text on identical captions shows that model differences affect cortical correspondence. It does not determine how much of the weak brain–model alignment arises because the brain receives images while the models receive incomplete textual descriptions.

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

This paper presents a tangent-space geometry framework for comparing fMRI responses in human visual cortex with Llama-3-8B caption representations. The paper offers a useful perspective by complementing voxel-wise ridge encoding with local tangent-space orientation over shared semantic neighborhoods, and it introduces a CLIP text-encoder comparison as a control. Overall, the paper is well motivated and the geometric diagnostic is useful, but the robustness of the central weak-alignment and layer-invariance claims is not sufficiently established. I consider this an interesting but currently insufficiently supported work.

Critical strengths
  1. The paper clearly distinguishes linear encoding prediction from local representational-geometry alignment and introduces tangent-space orientation as a complementary diagnostic, which is a conceptually useful contribution to the brain–model alignment literature.
  2. The analysis is systematic and multi-faceted, combining local variance spectra, tangent-space chordal distances, voxel-wise ridge encoding, a CLIP comparison, and steering-vector analyses across eight subjects with group-level statistics.
  3. The CLIP text-encoder comparison provides a meaningful control that reveals a qualitatively different depth-dependent correspondence pattern, showing that the layer-invariance observed for Llama is not a generic property of caption representations.
Critical weaknesses
  1. My primary concern is the absence of a noise ceiling for tangent-space alignment. The observed chordal distances are only slightly below the random baseline, and without a split-half or simulated-noise upper bound, it is impossible to determine whether the weak alignment is genuine or largely due to fMRI noise and local PCA estimation error.
  2. The semantic neighborhoods are defined using CLIP image embeddings, and the CLIP comparison uses a CLIP text encoder, so the two share the CLIP training objective. The depth-dependent CLIP effect may therefore partly reflect this shared structure, and no non-CLIP neighborhood replication is provided to test robustness.
  3. The choices d=8 and k=100 lack sensitivity analysis, multiple-comparison correction across ROIs and layers is unclear, and the V1 negative control in the steering intervention is not clean, so the current protocol does not support a causal interpretation.
Questions for authors
  1. Can the authors report a noise ceiling or split-half upper bound for tangent-space alignment and state the observed alignment as a proportion of that upper bound?
  2. Can the authors replicate the EBA > FBA > early-visual gradient and the CLIP depth effect using at least one non-CLIP semantic neighborhood definition? If the results are not robust, how would the authors interpret the central claim?
  3. Have the authors performed sensitivity analyses for d=8 and k=100?
  4. Could Llama’s reported layer-invariance partly stem from token mean-pooling and caption averaging, and would layer-to-layer variation remain as small under token-level or alternative pooling?
4 comments

We sincerely thank our reviewer for their valuable feedback and suggestions. We appreciate that our reviewer recognizes the contribution of our work to the brain-model alignment literature and for acknowledging the value of our multi-scale analysis. Our additional experiments demonstrate robustness and establish our central weak-alignment claims.

W2.1/Q2.1: To address the reviewer's concerns about measurement reliability, we performed a split-half analysis of cortical tangent-space geometry. We divided the repeated measurements into two subsets (T1 and the average of T2 and T3), independently estimated cortical tangent subspaces, and computed squared chordal distances within and across splits (Appendix Figure S1).

We use split-half cortical correspondence as an empirical reliability reference for brain--LLM alignment, rather than a formal noise ceiling, and define Faligned=Drandom−Dbrain,LLMDrandom−DsplitF_{\mathrm{aligned}} = \frac{D_{\mathrm{random}} - D_{\mathrm{brain,LLM}}}{D_{\mathrm{random}} - D_{\mathrm{split}}}. Here, Faligned=0F_{\mathrm{aligned}} = 0 corresponds to the random-subspace baseline, while Faligned=1F_{\mathrm{aligned}} = 1 indicates that brain-LLM distance equals the split-half cortical distance, without implying an upper bound on alignment. We summarize voxel counts, mean brain-LLM chordal distances across Llama layers, split-half cortical distances (T1+T2 vs. T3), and FalignedF_{\mathrm{aligned}}, using a random-subspace baseline of 7.354 for NSD Subject 1:

ROIVoxelsObserved (mean across layers)Split-halfFalignedF_{\mathrm{aligned}} (%)
V113507.245.948.1
V214337.256.5713.3
V311877.246.2410.2
hV46877.246.6716.7
EBA22386.886.5357.8
FBA6967.077.0490.4

Noise-normalized brain-LLM alignment was substantially higher in EBA (57.8%) than in V1 (8.1%), consistent with stronger geometric correspondence in body-selective cortex. FBA exhibited an even higher normalized value (90.4%), but this estimate is less reliable because its split-half distance is close to the random-subspace baseline. We therefore interpret the FBA result cautiously, without treating it as evidence of near-ceiling alignment.

W2.2/Q2.2 We appreciate our reviewer’s suggestion to assess whether our findings generalize beyond the CLIP-based semantic neighborhood definition. We repeated all neighbourhood-based analyses in four subjects (S1, S2, S5, S7) with neighbourhoods defined by SigLIP2 (so400m-patch14-224) image embeddings, keeping every other step identical (k=100k = 100, d=8d = 8, same ROIs, layers and preprocessing). SigLIP2 and CLIP neighbourhoods share only 52–53% of their members, so this is a substantial change. The main conclusions hold. Brain-Llama alignment remains close to the random-subspace expectation (7.354) and is strongest in EBA, then FBA, then early visual cortex in every subject (four-subject means with SigLIP2: EBA 6.886, FBA 7.067, V1–hV4 7.228; with CLIP: 6.936, 7.107, 7.237), and remains approximately constant across Llama layers. SigLIP2 neighbourhoods slightly strengthen alignment, mainly in body-selective cortex, and the CLIP text encoder’s advantage over Llama is preserved. The depth-dependent alignment of the CLIP text encoder is unchanged under SigLIP2 neighbourhoods (distance decreases with depth in 24/24 ROI × subject cases under both). It therefore does not stem from the shared CLIP objective. Within-system organisation (cortical ρ = 0.79, Llama ρ = 0.76) and local variance spectra are unchanged. We have incorporated this replication into the revised manuscript in Appendix S3.

Tschannen, M., Gritsenko, A., Zhai, X., et al. (2025). SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features. arXiv preprint arXiv:2502.14786. https://arxiv.org/abs/2502.14786

W2.3: We thank our reviewer for highlighting the importance of a sensitivity analysis for local neighborhood size kk and subspace dimensionality dd and carefully interpreting the results of our steering intervention. To address this concern, we tested the robustness of tangent-orientation geometry across 28 combinations of different neighborhood sizes (k \in{50, 100, 200, 400\}) and number of leading left singular vectors used to define each local variation subspace (d∈{2,4,6,8,12,16,20}d \in \{2, 4, 6, 8, 12, 16, 20\}) for NSD Subject 1 (Appendix S2). We report squared chordal distances normalized to the random-subspace baseline for each configuration (0 = perfect alignment; 1 = random-level dissimilarity) in Table S2 and Figure S2.

Across these tested configurations, the brain-model subspace overlap was limited, the region of interest (ROI) ordering was preserved, and the overlap profile across Llama layers remained largely flat for the mean-pooled token embeddings, which indicate that our findings are not specific to the original choice of k=100k=100 and d=8d=8. Pairwise Spearman rank correlations showed stable dissimilarity patterns for each of the blocks: structural consistency was highest for the brain-brain block median (ρ=0.982\rho = 0.982, range: 0.8710.871–1.0001.000), followed by the LLM-LLM block median (ρ=0.941\rho = 0.941, range: 0.7090.709–0.9990.999), and the brain-LLM block (ρ=0.771\rho = 0.771, range: 0.4650.465–0.9960.996). EBA and FBA showed lower brain-LLM dissimilarities than early visual cortex, with mean normalized distance differences ranging from −0.0544-0.0544 to −0.0210-0.0210 (EBA/FBA minus early visual cortex), suggesting that the relative cortical ordering is robust to the choice of parameters.

Regarding multiple-comparison correction across ROIs and model layers, our study quantifies brain-model alignment using chordal distances and describes observed region and layer-dependent patterns without formal hypothesis testing. Therefore, multiple-comparison correction is not applicable.

Q2.4: We thank our reviewer for suggesting that we examine whether the observed layer-invariance in brain-Llama alignment persists under alternative representation extraction strategies. We agree that our representation aggregation strategy -- averaging token embeddings for each caption, and subsequently averaging over captions for each image -- may diminish layer-specific variation, leading to the largely flat alignment profile observed across Llama layers. To assess the effect of mean token pooling, we repeated the ridge-encoding for NSD Subject 1 using last-token representations instead of mean-pooled caption embeddings, and compare both brain-LLM chordal dissimilarities and cross-pooling dissimilarities between mean-pooled and last-token Llama representations (Appendix S4).

Last-token representations yielded lower brain-encoding performance than mean pooling but showed greater variation across Llama layers (maximum within-ROI difference in Pearson r=0.012r = 0.012 in mean-pool vs 0.0370.037 in last-token) , with performance generally improving with depth (Appendix Figure S10). Chordal distances between Llama layers for the same representation choice increased with layer separation in both last-token and mean-pooling, but were smaller with mean pooling (D(L4,L32)=3.00D(L4, L32) = 3.00 in mean-pool vs 4.284.28 in last-token) (Appendix Figure S11). Comparisons across pooling strategies showed substantial dissimilarity in their local neighborhood tangent-space orientations (4.604.60–6.116.11), even between corresponding layers. Brain-LLM chordal distances were relatively lower and more stable across layers under mean pooling; last-token distances were relatively higher and generally decreased with depth. Under mean pooling, distances ranged from 6.876.87–6.916.91 in EBA, 7.077.07–7.107.10 in FBA, and 7.247.24–7.277.27 in early visual cortex; under last-token pooling, the corresponding ranges were 7.107.10–7.267.26, 7.197.19–7.297.29, and 7.297.29–7.357.35. Token pooling affects variation across Llama layers and their representational orientations, cross-system geometric findings are preserved, with the lowest brain-LLM distances for any LLM layer in EBA followed by FBA and early visual cortex. We have also revised the main text to clarify that the apparent invariance across Llama layers depends on token pooling.

Thank you for the detailed response and the additional analyses. The new results address several of my concerns about neighborhood choice, parameter sensitivity, and token pooling. As you show in your response to W2.1, even repeated fMRI measurements of the same brain region can yield weakly aligned local subspaces, particularly in FBA. This reinforces my concern about how much of the observed weak brain–Llama alignment may be due to fMRI measurement noise and variability across repetitions. For that reason, I’m keeping my original rating of 2 (weak reject). I also noticed that the FBA value reported in the text of Appendix S1 (88.9%, around line 808) appears to differ from the value in Table S1 (90.4%), so I’d appreciate it if you could double-check the reported numbers throughout the manuscript.

We thank our reviewer for identifying this discrepancy. The correct FBA value is 90.4%, as reported in Table S1. The value of 88.9% in the Appendix S1 text resulted from a rounding inconsistency in the underlying chordal distance used for normalization. We have corrected the Appendix S1 text to 90.4% to match Table S1 and checked the reported values throughout the manuscript for consistency.

Furthermore, we have clarified the distinction between descriptive analyses of regional and layer-dependent brain–model alignment and formal statistical hypothesis tests. The regional and layer-dependent patterns are presented descriptively, without claims of statistically significant differences between individual ROIs or model layers. For the three one-sample t-tests reported in Section 4.3, which compare cross-system chordal distances against the random-subspace baseline, we have applied Holm–Bonferroni correction across the three ROI groups. All three comparisons remain significant after correction, and we have updated the manuscript to report the adjusted p-values.

Reviewer_85fmOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper compares linear encoding and local representational geometry between Llama caption embeddings and visual cortical responses from eight NSD participants. Llama predicts cortical responses, but the dominant local variation subspaces show limited overlap and little variation across layers. A CLIP text encoder shows increasing brain correspondence with depth.

Critical strengths

The distinction between linear predictability and local geometric similarity is useful. The regional differences replicate across eight participants. Encoding is evaluated on held-out images, and the random-subspace baseline accounts for centering. The authors appropriately describe the intervention results as exploratory.

Critical weaknesses
  1. The interpretation of limited overlap needs a reliability reference. Section 4.3 acknowledges that no noise ceiling is estimated. Without assessing the reproducibility of cortical local subspaces across repeated measurements, it is difficult to determine how much of the limited brain-model overlap reflects representational differences rather than fMRI noise. The random baseline establishes correspondence above chance, but does not indicate its strength relative to what can be reliably measured. This does not invalidate the observed distances, but limits their interpretation as evidence of stable geometric differences. A repeated-trial reliability analysis or a positive control with matched measurement noise would help address this concern.

  2. The main geometric results are reported for one neighborhood size and subspace dimension. The analysis uses k = 100 and d = 8, while the variance spectra differ substantially across Llama layers. The paper does not establish whether limited overlap, the regional ordering, and the largely flat depth profile persist under other reasonable choices of k and d. A small sensitivity analysis, comparing effects relative to the corresponding random baselines, would help establish that these findings do not depend on the selected analysis scale.

Questions for authors

Please refer to the Weakness part

3 comments

We appreciate our reviewer’s recognition that the linear-predictability-vs-geometric-similarity distinction is useful and that the regional differences replicate across all eight participants. We address both weaknesses below.

W1.1: We agree that a reliability reference is necessary to distinguish measurement-related variability from genuine representational differences, and that the random-subspace baseline alone does not account for fMRI measurement variability. To address this concern, we add a repeated-trial analysis for Subject 1, splitting the three measurements per stimulus into one trial (T1) and the average of the other two (T2/T3), independently estimating cortical tangent subspaces, and comparing squared chordal distances within and across splits (Appendix Figure S1).

The relative organization of cortical ROIs was reproducible across splits and consistent with the within-cortex gradient reported in the main analysis. Matched-ROI distances ranged from 5.94 (V1) to 7.04 (FBA), compared with a random-subspace expectation of 7.354, indicating considerable variability in tangent-space orientations.

We use split-half cortical correspondence as an empirical reliability reference for interpreting brain-LLM alignment, rather than as a formal noise ceiling. We quantify alignment relative to this reference as: Faligned=Drandom−Dbrain,LLMDrandom−Dsplit.F_{\mathrm{aligned}} = \frac{D_{\mathrm{random}} - D_{\mathrm{brain,LLM}}}{D_{\mathrm{random}} - D_{\mathrm{split}}}.

Here, Faligned=0F_{\mathrm{aligned}} = 0 indicates alignment at the random-subspace baseline, while Faligned=1F_{\mathrm{aligned}} = 1 indicates that the brain--LLM chordal distance equals the split-half cortical distance. Importantly, Faligned=1F_{\mathrm{aligned}} = 1 is a descriptive reference point, not a theoretical upper bound on achievable alignment. For Subject 1, we summarize voxel counts, mean brain--LLM and split-half cortical distances, and FalignedF_{\mathrm{aligned}} (random baseline: 7.354). ROI--layer results appear in Table S1.

ROIVoxelsObserved (mean across layers)Split-halfFaligned (%)
V113507.245.948.1
V214337.256.5713.3
V311877.246.2410.2
hV46877.246.6716.7
EBA22386.886.5357.8
FBA6967.077.0490.4

Noise-normalized brain-LLM alignment was substantially higher in EBA (57.8%) than in V1 (8.1%), consistent with stronger geometric correspondence in body-selective cortex. Although FBA showed a higher normalized value (90.4%), this estimate may be unstable because its split-half distance approaches the random-subspace baseline. We therefore interpret FBA cautiously rather than as evidence of near-ceiling alignment.

W1.2: We also thank our reviewer for highlighting the importance of evaluating sensitivity of our geometric results to local neighborhood size kk and subspace dimensionality dd. To address this concern, we conducted a sensitivity analysis across different neighborhood sizes (k \in{ 50, 100, 200, 400\}) and on the number of leading left singular vectors used to define each local variation subspace (d∈{2,4,6,8,12,16,20}d \in \{2, 4, 6, 8, 12, 16, 20\}) for NSD Subject 1. Across the tested configurations, the chordal dissimilarity matrix structure remained incredibly consistent: the brain-model subspace overlap was limited, the region of interest (ROI) ordering was preserved, and the overlap profile across Llama layers remained largely flat. We provide the joint brain x LLM tangent-orientation dissimilarity matrices normalized relative to the random subspace values for each configuration of k and d (Table S2, Figure S2). A normalized value of 0 indicates perfect alignment, whereas 1 indicates random-level alignment.

To quantify the stability of these patterns, we computed Spearman rank correlations between the corresponding blocks (brain-brain, LLM-LLM, brain-LLM) across all pairs of parameter configurations. Structural consistency was highest for the brain-brain block median (\rho = 0.982, range = 0.871-1.000), followed by the LLM-LLM block median (ρ=0.941\rho = 0.941, range: 0.7090.709–0.9990.999, and the brain–LLM block (ρ=0.771\rho = 0.771, range: 0.4650.465–0.9960.996). The EBA/FBA versus early visual cortex contrast was preserved across all 28 configurations, with EBA and FBA consistently exhibiting lower normalized brain–LLM dissimilarities (contrast range: −0.0544-0.0544 to −0.0210-0.0210). These results, included in Appendix S2, demonstrate that our geometric findings are robust to reasonable variations in neighborhood size and subspace dimensionality.

Thank you for the additional analyses. The kk, dd sensitivity analysis addresses my second concern. My reliability concern remains partly unresolved. Appendix S1 shows a reproducible pattern across brain regions, alongside substantial variability in subspace estimates across repetitions. The split-half ratio provides useful context, but does not by itself establish how much of the low brain–LLM overlap comes from measurement noise rather than stable representational differences. A positive control with noise matched to the main analysis would help distinguish these explanations.

Figures S8 and S15 acknowledge a confound involving the first token and caption length in the mean-pooled PCA spectra. However, these spectra remain part of the main findings without a targeted test of how much this confound contributes to the reported depth trend. The last-token comparison changes the aggregation strategy more broadly and therefore does not isolate the first token’s contribution. Keeping mean pooling while excluding the first token would provide a more targeted control.

There is also an inconsistency in the reply to Reviewer wvdW: it explains that multiple-comparison correction is inapplicable because the study uses no formal hypothesis testing, yet Section 4.3 reports t-tests and p-values for three ROI groups. Please clarify the scope of that statement and whether those p-values were corrected. This reporting issue is separate from the substantive concerns above, which still limit the conclusions I can draw. I will keep my score.

We thank our reviewer for identifying this inconsistency. The statement in our response to Reviewer wvdW (W2.3) was imprecise. Section 4.3 reports three one-sample t-tests comparing the subject-level mean cross-system chordal distances against the random-subspace baseline (7.354) for the three ROI groups: V1–hV4, FBA, and EBA (8 subjects, df=7df=7). The uncorrected p-values were 3.59×10−63.59\times10^{-6}, 1.28×10−51.28\times10^{-5}, and 2.10×10−52.10\times10^{-5}, respectively.

We applied Holm–Bonferroni correction across these three comparisons at α=0.05\alpha=0.05. All three comparisons remained significant after correction (V1–hV4: pHolm=1.08×10−5p_{\mathrm{Holm}}=1.08\times10^{-5}; FBA: pHolm=2.55×10−5p_{\mathrm{Holm}}=2.55\times10^{-5}; EBA: pHolm=2.55×10−5p_{\mathrm{Holm}}=2.55\times10^{-5}). We have updated Section 4.3 to report the corrected p-values. The conclusion that all three ROI groups exhibit mean cross-system chordal distances significantly below the random-subspace baseline remains unchanged.

Comments on the paper

1 comment

We sincerely thank all three reviewers for their detailed and constructive reviews. All reviewers appreciated the main thesis of the paper. Reviewer 85fm notes that "the distinction between linear predictability and local geometric similarity is useful"; Reviewer wvdW identifies the distinction between "linear encoding prediction" and "local representational-geometry alignment" via "tangent-space orientation as a complementary diagnostic" as a "conceptually useful contribution to the brain-model alignment literature", and considers the paper “well motivated and the geometric diagnostic useful”; Reviewer HZkm notes that the paper "presents a new perspective for the alignment of the LLM and brain representation" and that preliminary results support that "some linear relation exists between LLM activation and brain stimulus patterns."

The reviews converge on a few methodological and conceptual questions. Reviewers 85fm and wvdW raise within-system reliability / noise-ceiling concern (85fm W1, wvdW W1/Q1), and request a (d, k) sensitivity analysis (85fm W2, wvdW Q3) . Reviewer wvdW asks for pooling robustness (Q4) and a non-CLIP semantic-neighborhood control (W2/Q2). Reviewer HZkm raises the conceptual point (W1/Q1) of the significance of the subspace alignment.

We conduct robustness analysis using repeat-trial reproducibility and ablations of neighborhood size kk and subspace dimension dd on Subject 1, the highest-SNR subject in NSD. We independently estimated cortical tangent subspaces from two independent splits (one containing one of three trials and the other containing the average of the remaining two). We observed the same pattern of ROI-graded alignment across the independent trial subsets, and a clear two-block structure ({V1, V2, V3, hV4} vs. {EBA, FBA}) in both within-split and cross-split chordal distance matrices, supporting the reproducibility of relative cortical organization.

In our sensitivity analysis, we show that the overall geometric structure and regional alignment contrasts remain substantially consistent across different choices of neighborhood size kk and subspace dimension dd. We extended the steering experiment from FBA at L24 vs. V1 at L4 to all four combinations (FBA/V1 × L4/L24) across four NSD subjects, showing a similar effect overall but non-uniform across categories. Switching from content-token mean pooling to last-token representations preserves the regional ordering of brain-LLM alignment, with the strongest correspondence in EBA, followed by FBA and early visual cortex. Last-token representations show depth dependence, with alignment generally increasing across deeper Llama layers, and weaker overall alignment.

To assess generalization, we extend the analysis to Qwen2.5-7B and the CLIP ViT-L/14 text encoder on NSD, and evaluate Llama-3-8B and CLIP text representations on two additional fMRI datasets, BOLD5000 and THINGS-fMRI (Appendices S5–S7 and S13). Both Llama-3-8B and Qwen2.5-7B show positive encoding accuracy and weak geometric alignment with visual cortex, with the strongest correspondence in EBA and FBA. We also test the robustness of our geometric results by defining semantic neighborhoods with SigLIP2 instead of CLIP (Appendix S3). These additional experiments further validate the scientific value of our results.

Per-reviewer replies address each question in order. The major revisions are marked in blue in the revised manuscript. For clarity of presentation, a new Appendix section is identified by a blue title.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.