- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper compares fMRI data from six Chinese-English bilingual participants who read several hours of narratives in both languages. Specifically, they compare spatial cortical organisation and semantic tuning of semantic representations across subjects. To make these comparisons, they use some common brain alignment techniques (voxelwise encoding models) and build upon the analysis of Chen et al. by doing a clustering of model weights with 5 centroids that they associate with categories and then basing their analysis on the association to these clusters. The findings are that the cortical overlap was higher across languages than among participants, while the semantic tuning similarity showed the opposite pattern. But they admit that these results depend on the mapping of the individual cortical surfaces to the common fsaverage space.
Critical strengths
The paper is well written and explained, and the questions are well motivated. The experiments are clear, and the procedure is correct. They use common brain alignment techniques, banded ridge regression, and temporal correction, and fMRI data that was already used in the literature.
Critical weaknesses
-
(W1) While ICLR welcomes computational neuroscience work, such as encoding models or RSA comparing model and brain representations, the machine learning angle of this paper is limited. The contribution is primarily neuroscientific since it compares the organisation of semantic representations across bilingual participants, and the rest of the pipeline (voxelwise encoding models, projection to a common surface space, and clustering of model weights) combines well-established techniques. The paper does not offer new insights into model representations or into the relationship between model and brain representations and does not frame its contribution to an ML audience.
-
(W2) The contribution appears incremental relative to prior work with an almost identical design and pipeline (Chen et al. 2026). The main difference is that, with the same data and the same methodology, once Chen et al. established cross-language consistency within each brain, this paper adds the cross-participant comparison; it now compares 15 pairs per language (30 total), using Pearson correlation on the re-averaged weight vectors for semantic tuning and the Dice coefficient for cortical overlap.
-
(W3) The paper's central claim is precisely that the cross-participant overlap is low. Comparing across subject brains required projecting all subjects onto a common space (fsaverage). In Chen et al., they also do this projection, but their comparisons were within the same brain, so the answer did not depend on the space. The problem is that imperfect individual-to-template mapping mechanically lowers the cross-participant Dice. The artefact and the expected result point in the same direction, so a low Dice cannot by itself support the conclusion. The authors acknowledge this in their own limitations section: “Comparisons of cortical organisation across participants rely on mapping individual cortical surfaces to the common fsaverage space. Imperfect alignment may reduce the measured cortical overlap across participants.”
Questions for authors
-
(W3) Run the experiments proposed in the future work to examine how robust the pattern found is to different alignment methods.
-
(W3) The limitation discusses how mapping individual cortical surfaces to fsaveragre may affect the cortical organisation results. Could the authors also discuss how this mapping may affect the semantic tuning results? This mapping, beyond being used to compute the centroids, is also used to determine the cluster assignment that results in the mean weight vector for the cross-participant comparisons (15 per language).
-
What is the pairing scheme in the paired t-tests on Figures 3 and 4?
Critical weaknesses and questions
We thank the reviewer for their thorough and constructive review. We appreciate that they found the paper well written and explained, and the question well motivated. We are encouraged that they found the experiments clear and the procedure correct.
Critical weaknesses: W1) We agree that our main contributions are neuroscientific. We have added suggestions for future work aimed at a machine learning audience. Lines 510-513 in the Future work section now read: “Our comparison across participants and languages motivates a parallel question for multilingual language models. How do semantic representations vary across independently trained instances of the same model when they process the same language, and how do they vary across languages within a single instance?”
W2) Our study advances beyond Chen et al. (2026) in three ways. First, we explicitly differentiate the cortical location of semantic representations from their semantic tuning. Second, we compare variation across participants with variation across languages, thereby placing bilingual differences in the context of individual differences. Third, this comparison reveals opposite patterns: semantic tuning is more similar across participants within a language, whereas cortical location is more consistent across languages within a participant. This shows that semantic tuning of representations can be shared across participants even when their cortical location is participant-specific. Chen et al.’s within-participant comparisons could not reveal this pattern.
Questions:
(W3) Run the experiments proposed in the future work to examine how robust the pattern found is to different alignment methods. We agree that imperfect alignment can reduce cross-participant cortical overlap. To assess whether the pattern holds under a different anatomical mapping, we repeated the analysis in MNI space instead of fsaverage. We used affine registration to map each participant’s normalized model weights and prediction scores to FSL’s 2 mm MNI152 T1 brain template. We then performed the clustering and the comparisons as in the original analysis. The results are reported in Appendix section A.4. All analyses were conducted in MNI space. The resulting clusters were projected onto fsaverage only for visualization on a single cortical map. For both mBERT and fastText, as in our original analysis, cortical overlap is significantly higher across languages within a participant than across participants within a language. And the opposite pattern is observed for semantic tuning similarity as well. Thus, we observe the same ordering in both the MNI and fsaverage template spaces. This suggests that our findings are not specific to the fsaverage projection. This does not eliminate alignment error as a contributing factor, and we have made this qualification more explicit in the Limitations.
Could the authors also discuss how this mapping may affect the semantic tuning results? Cluster assignment is based entirely on model weight vectors in embedding space, without using cortical coordinates or requiring vertices assigned to the same cluster to occupy corresponding locations across participants. Centroid estimation depends on anatomical alignment, which affects how weights are averaged across participants at corresponding template vertices and which vertices pass the group-average prediction threshold in both languages. We verified that assigning clusters after projecting weights into fsaverage gives the same assignments as assigning clusters in native space and then projecting the labels into fsaverage.
What is the pairing scheme in the paired t-tests on Figures 3 and 4? For both cortical overlap and semantic tuning similarity, we computed one cross-language average and one cross-participant average per participant, then compared these six pairs using a paired t-test.