Clarification of fine-tuning methodology

We thank the reviewer for their valuable comment. We would like to highlight that the novelty of our work lies in the design of the brain-informed fine-tuning pipeline. Our pipeline is a general, end-to-end framework that addresses key limitations of prior brain-to-model alignment methods and enables direct integration of brain data into language model fine-tuning.

Most previous approaches aligned stimulus features to brain responses via preprocessing steps based on rigid assumptions; for instance, using fixed word presentation rates or downsampling (e.g., Schwartz et al., 2019). These designs impose strong constraints. In contrast, our pipeline allows for:

• variable-length input (since word timing per repetition time (TR) can vary),
• more flexible and generalizable alignment applicable across modalities and tasks as we pass raw text as input
• dynamic stimulus-repetition time (TR) alignment as part of the learning process, not a static preprocessing step

To illustrate this advantage, we implemented the preprocessing-based alignment (as in prior work) and fine-tuned a new model on bilingual data from one participant. The mean encoding performance is 0.13 for the preprocessing-based method, compared to 0.16 for our brain-informed fine-tuned model. We hope this highlights the practical and conceptual contributions of our approach. We will add this reasoning in the camera-ready to ensure clarity.

Clarification on gains in Table 1

Thank you for raising this point. The initial Table 1 reported downstream task performance for a single participant, which may have appeared inconsistent. We have revised Table 1 to show average performance across all six bilingual participants, with standard deviation. The overall trends remain consistent with those initially reported, helping strengthen the robustness of the findings. Due to space constraints, please see the updated table 1 included in our response to reviewer hVbZ. This aggregated analysis strengthens the robustness and practical utility of our approach, showing consistent benefits across participants and models. Please note that models were fine-tuned separately for each participant and language.

Q1

Thank you for the question. We placed the results in the appendix due to space constraints, not due to lack of performance. In fact, all three models (mBERT, XLM-R, and XGLM) show improvements (see below) following brain-informed fine-tuning. This table shows that fine-tuning improves performance on most tasks: mBERT (7/9 tasks), XLM-R (9/9), and XGLM (6/9), with especially large gains for XLM-R (e.g., RTE: 54.15 → 62.71, MRPC: 69.61 → 77.94).

Updated Supplementary Table 2. a) Fine-tuning and Evaluation in the Same Language

Q2

Thank you for the helpful feedback. We acknowledge that Figure 2 has a lot of information. The main takeaway from Figure 2 is that across both languages and model types (monolingual and multilingual), bilingual brain-informed fine-tuning consistently yields better encoding performance than the vanilla models in several brain regions.

While spatial patterns vary due to inter-individual differences (see Appendix Figure 5), we observe improvements in encoding performance in several high-level language areas.. Given this variability, the improvements in downstream task performance from fine-tuning are still robust. We will provide separate flatmaps for each fine-tuning condition in the supplementary material to improve clarity.

Q3

Thank you for this important question. Our work demonstrates that brain-informed fine-tuning can reliably enhance multilingual language model performance, even with limited data. Please note that each model was fine-tuned using 2.5 hours of naturalistic brain data per participant. We adopt a small-N design (Smith & Little, 2018), conducting full replication across participants by fine-tuning and evaluating models individually. While the performance gains may appear modest, they are robust across participants. Our results contribute to work on aligning brain and language model representations. To our knowledge, this is the first study to leverage bilingual brain data for fine-tuning, showing gains in multilingual understanding. This work offers a step toward bridging neuroscience and NLP to interpret and model multilingual semantic representations.

Q1.1

Thank you for this question. We focused primarily on encoder-based models (BERT, mBERT) for two reasons:

• Consistency with prior work: Previous brain-informed tuning approaches with monolingual english brain data have predominantly used encoders (e.g., BERT, Wav2Vec, Whisper, WavLM) due to their strong alignment with text and speech features.
• Bilingual dataset requirements: Our dataset contains bilingual brain recordings. To ensure architectural uniformity across English and Chinese experiments, we used BERT-base, BERT-Chinese, and mBERT, which share a comparable encoder backbone.

However, our experiments are not limited to encoders. We also fine-tuned XLM-R (encoder) and XGLM (decoder), showing that our brain-informed fine-tuning approach generalizes across architectures (see Supplementary Table 2). This table shows that fine-tuning improves performance on most tasks: mBERT (7/9 tasks), XLM-R (9/9), and XGLM (6/9), with especially large gains for XLM-R (e.g., RTE: 54.15 → 62.71, MRPC: 69.61 → 77.94).

Updated Supplementary Table 2. a) Fine-tuning and Evaluation in the Same Language

Q1.2

When applying brain-tuning to decoder-only models such as XGLM and LLaMA, we observed two challenges:

• data efficiency; these larger generative models require more neural data and are prone to overfitting
• task mismatch; our downstream evaluations focus on understanding rather than generation.

Despite this, we observed small gains, suggesting that the method is architecture-agnostic but currently more effective on encoder models given data and task constraints.

Please see below for results with LLaMA‑3.2‑1B fine-tuned using our brain-informed approach with LoRA adaptation applied to layer 9 (similar to Antonello et al., 2024). Models were evaluated on three downstream datasets (Wang et al. 2019, Sakaguchi et al., 2019). Brain-informed fine-tuning yields improvements over the base model across two datasets, further supporting the generalizability of our approach to different architectures and low-parameter adaptation methods (LoRA). We note that improvements are small, likely due to the parameter-efficient adaptation (LoRA on a single layer). Unfortunately, we couldn’t perform full fine-tuning on this model due to computational constraints.

Q2

In response to your suggestion, we implemented an alternative version of our pipeline that mirrors the preprocessing-based approach used in prior work; stimulus-TR alignment is performed prior to fine-tuning. Specifically, we constructed stimulus-TR pairs similar to Schwartz et al. (2019). We then fine-tuned the language model using this pre-aligned input and bilingual brain data from 1 participant. The mean encoding performance is 0.13 for the preprocessing-based method, compared to 0.15 for the vanilla model and 0.16 for our fine-tuned model. The table below presents downstream task performance (on the GLUE benchmark) for this preprocessing-based method (denoted as BERT-ft-en (Ablation)). This allows direct comparison with our proposed end-to-end brain-informed fine-tuning pipeline. We observe a substantial drop in performance for the preprocessing-based method (BERT-ft-en (Ablation)). This suggests that stimulus-TR alignment performed prior to fine-tuning procedure likely discards valuable temporal and contextual information. In contrast, our approach integrates brain alignment during fine-tuning, enabling robust downstream performance.

Q3

Thank you for this thoughtful question. We analyzed scaling behavior on high-performing voxels using up to 6 stories (~300 TRs each). While there is a mild upward trend in encoding performance (mean r: 0.151 → 0.162), the effect is small relative to the standard errors, likely due to the limited dataset size. We expect clearer scaling trends to emerge with larger brain datasets (Antonello et al., 2023).

[Antonello et al., 2023] Scaling laws for language encoding models in fMRI

Q4

Thank you for this interesting question. We observe that language-selective and semantic-selective voxels provide stronger gains in downstream task performance compared to using whole-brain voxels, suggesting that brain-informed fine-tuning primarily enhances semantic representations. Please see Appendix Table 5 for details. Our current work focused on improving encoding and task performance of bilingual brain-informed fine-tuning. However, we agree that interpreting which specific aspects of language (semantics, syntax, etc.) are altered is an important next step. We plan to explore attribution-based analyses and causal graph methods (Lindsey et al. 2025) in future work to better understand changes introduced with fine-tuning.

[Lindsey et al. 2025] On the Biology of a Large Language Model

With the extra page allowance, we will move key results from the appendix into the main text to improve clarity as suggested.

Inclusion of control task

Thank you for the helpful suggestion. We ran a control experiment on the GSM8K dataset (arithmetic reasoning task), and as expected found no performance gains; both vanilla and fine-tuned models achieved 0 accuracy and F1.

p.4 Clarification on usage of last hidden layer for the representation

Thank you for this important question. We think the confusion arises from the assumption that both analyses, brain-informed fine-tuning and voxel-wise encoding have the same goal, i.e., predicting brain activity from pretrained LLM representations.

This is not the case. We have three scenarious in our analysis. First, in brain-informed fine-tuning (Section 3.3) our aim is to induce a brain-informed bias into the language model’s representations by performing full end-to-end fine-tuning of all layers using brain data, without any layer selection. Second, we use the representations from the last hidden layer,following standard practice for downstream NLP tasks. Third, Once the model is fine-tuned, our goal is to evaluate how these brain-tuned language model representations perform, particularly in how well they predict brain activity. In this third case, we follow established provedures from the literature and focus on intermediate layers (e.g., layer 7 for BERT). We will revise the section to make this explicit and ensure clarity.

p.4 Comparison of different losses

Thank you for this important point, and apologies for the oversight regarding the missing hybrid loss. Please see the updated Supplementary Table 1 below for a full side-by-side comparison of all five loss functions, including the newly added hybrid loss, for mBERT-ft-en across GLUE tasks.

Updated Supplementary Table 1 a) Fine-tuning and Evaluation in the Same Language

We agree that Ridge loss achieves the strongest downstream performance on several tasks. We would like to note that contrastive loss yielded the better brain encoding performance (mean r = 0.163, SE = 0.0097), slightly outperforming Ridge (mean r = 0.1622) and MSE (mean r = 0.1612). Given that contrastive objectives are commonly used in prior brain decoding studies and offer strong alignment with brain representations (Défossez et al., 2023; Levy et al., 2025), we retained it as our main objective to ensure comparability with existing literature.

We will revise the manuscript to clarify this rationale and include both within-language and cross-language results for all five loss functions (including the hybrid loss) in the Appendix to provide a comprehensive side-by-side comparison.

[Défossez et al., 2023] Decoding speech perception from non-invasive brain recordings

[Levy et al., 2025] Brain-to-Text Decoding: A Non-invasive Approach via Typing

p.6 Clarification on improvements in visual areas

Thank you for this thoughtful observation. We agree that improvements in visual areas warrant further discussion. Please note this is a reading experiment, where variation in visual brain areas is correlated with stimulus presentation, especially in naturalistic settings. In addition, semantic features from language models are known to spuriously predict brain activity in visual areas even after regressing out low-level visual information (Oota et al., 2024; Deniz et al., 2019). Importantly, our analyses using semantically selective voxels (which explicitly exclude visual regions) show consistent performance gains from brain-informed fine-tuning. This indicates that the improvements are not driven by visual areas but reflect genuine alignment with higher-level semantic processing. We will add a discussion of this point in the revised manuscript.

[Oota et al., 2024] Speech language models lack important brain relevant semantics

[Deniz et al., 2019] The Representation of Semantic Information Across Human Cerebral Cortex During Listening Versus Reading Is Invariant to Stimulus Modality

p.6 Comparative statistics for control baselines

Thank you for the helpful suggestion. We will include visualizations of differences in encoding performance in the brain to better highlight improvements from brain-informed fine-tuning over controls. Additionally, we’ll provide descriptive statistics, including mean Δr per subject, across both baselines, models, and cross-participant analyses in the camera-ready version. For now, here is a summary statistic for BERT-en, across participants (for well-predicted voxels; r>0.1) to aid interpretation:

• Δr̄ (vanilla – fine-tuned): -0.037 ± 0.00037
• Δr̄(vanilla – participant transfer): -0.00055 ± 0.00029
• Δr̄(vanilla – control (TR shuffle)): 0.133
• Δr̄(vanilla – control (mBERT)): 0.136

As can be seen from the summary statistics reported above, our results remain consistent taking Δr in account.

p.7 Clarification on WNLI task

Thank you for pointing this out. The identical score (56.34) across some fine-tuned models on WNLI is due to the small dataset size and integer-valued confusion matrices (please refer to GLUEBenchmark FAQ). Note that our fine-tuned models do yield slightly different predictions, but the resulting accuracy rounds to the same value. For example, here are the confusion matrices for three different fine-tuned models (for 1 participant):

• BERT-ft-en: [[33, 18], [13, 7]]
• mBERT-ft-en: [[32, 17], [14, 8]]
• mBERT-ft-zh: [[34, 19], [12, 6]] We've updated Table 1 to report mean ± standard deviation across participants to better highlight subtle differences and clarify variability.

p.7 Comparison across participants and models

Thank you for the helpful feedback. We have now revised Table 1 to report the average performance across all six bilingual participants, along with standard deviations. All models were fine-tuned separately for each participant and language. The overall trends remain consistent with those initially reported, helping strengthen the robustness of the findings. Due to space constraints, please see the updated supplementary table 2 included in our response to reviewer xVwE. Regarding model comparisons, the current supplementary material includes fine-tuned results for mBERT, XLM-R, and XGLM. We apologize for the oversight in not including the performance of their corresponding vanilla models and LLaMA-3.2-1B. Please see the updated table below. This table shows that fine-tuning improves performance on most tasks: mBERT (7/9 tasks), XLM-R (9/9), and XGLM (6/9), with especially large gains for XLM-R (e.g., RTE: 54.15 → 62.71, MRPC: 69.61 → 77.94).

Please see below for results with LLaMA‑3.2‑1B fine-tuned using our brain-informed approach with LoRA adaptation applied to layer 9 (similar to Antonello et al., 2024). Models were evaluated on three downstream datasets (Wang et al. 2019, Sakaguchi et al., 2019). Brain-informed fine-tuning yields improvements over the base model across two datasets, further supporting the generalizability of our approach to different architectures and low-parameter adaptation methods (LoRA). We note that improvements are small, likely due to the parameter-efficient adaptation (LoRA on a single layer). Unfortunately, we couldn’t perform full fine-tuning on this model due to computational constraints

We will update the supplementary section to include these results for completeness.

[Wang et al. 2019] SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems, NeurIPS-2019

[Sakaguchi et al. 2019] WinoGrande: An Adversarial Winograd Schema Challenge at Scale, 2019

p.7 Clarification and extension of bilingual vs monolingual brain data

Thank you for pointing this out. To address the reviewer’s concern about robustness, we increased the number of monolingual participants by adding two additional participants from another dataset (LeBel et al., 2023). Due to space constraints, please see this extended comparison with 3 monolingual participants (updated table 2) included in our response to reviewer hVbZ. The results show that bilingual brain fine-tuning consistently outperforms the mBERT baseline across tasks and offers slightly stronger generalization than fine-tuning on individual monolingual participants. Bilingual brain fine-tuning outperforms monolingual fine-tuning on 7 out of 8 tasks on the CLUE benchmark, with equal performance on the remaining task (ChID). We will correct the misleading wording and be more explicit in the final version.

[LeBel et al., 2023] A natural language fMRI dataset for voxelwise encoding models

Minor point; p.4: We use a dropout of 0.2 for brain-informed fine-tuning to mitigate overfitting on limited brain data, and 0.1 for downstream tasks, following standard BERT fine-tuning defaults. We will include this information in the camera ready version.

Minor point; Fig. 2 p.6 and Table 5 p.26: Thanks for catching this. We’ll correct the appendix reference and reorder tasks accordingly.

Major comments

Q1 sample size

Thank you for raising this important concern. Our study shows that brain-informed fine-tuning can reliably enhance multilingual understanding. We employ a small-N design, where a language model is fine-tuned and evaluated independently for each participant, enabling robust within-subject inference (Smith & Little, 2018). Importantly, each participant in our dataset contributes two hours of naturalistic brain recordings, which are critical for capturing the complexity of language representations (Hamilton & Huth, 2018) in the brain and for enabling meaningful learning through fine-tuning. While we recognize that larger sample sizes would strengthen the generalizability of our findings, data collection is not feasible at this stage due to current resource constraints (four 2.5 hours sessions for data collection, data analysis, preprocessing and subject recruitment time). That said, we fully agree with the reviewer’s suggestion and currently plan to expand the bilingual dataset and extend our main findings to new participants in future work.

We have increased the number of monolingual participants, please see response to Q4 below.

[Smith & Little, 2018] Small is beautiful: In defense of the small-N design

[Hamilton & Huth, 2018] The revolution will not be controlled: natural stimuli in speech neuroscience

Q2 Reporting variability

Models were fine-tuned separately for each participant and language (note Methods 3.3). We have now revised Table 1 to report the average performance across all six bilingual participants, along with standard deviations. As shown in the updated results (see below), the results remain consistent with those originally reported in the paper. We thank the reviewer for this helpful feedback as this change should help better demonstrate the robustness of our findings.

Updated Table 1, a) Fine-tuning and Evaluation in the Same Language

Updated Table 1, b) Cross-Language Transfer Between Known Languages

Updated Table 1, c) Zero-Shot Transfer to Unseen Languages

Q3 Clarity in Table 1

We understand how this may have caused confusion. To maintain consistency across tasks, we followed the common practice used in the GLUE leaderboard and reported all metrics on a 0-100 scale. For example, a Pearson correlation of 0.88 is reported as 88.0. To improve clarity, we will add metrics adjacent to the task (like in the table above) and a footnote to all relevant tables explicitly indicating the evaluation metric range and the scaling.

For reference, please see the table below:

Q4 Monolingual participant experiments

In response to your feedback, we identified two additional monolingual participants from another dataset (LeBel et. al, 2023) and have now included them in the analysis. With this expanded comparison group (3 monolingual participants), the results show that bilingual brain fine-tuning consistently outperforms the mBERT baseline across tasks and offers slightly stronger generalization than fine-tuning on individual monolingual participants. Bilingual brain fine-tuning outperforms monolingual fine-tuning on 7 out of 8 tasks on the CLUE benchmark, with equal performance on the remaining task (ChID). We thank the reviewer for this suggestion as we think this extended analysis better supports the contribution of our paper and makes our results more convincing.

Updated Table 2. b) Cross-Language Transfer Between Known Languages

[LeBel et al., 2023] A natural language fMRI dataset for voxelwise encoding models

Q5 Clarification on brain region analysis

The reviewer is correct that Tables 1 and 2 are based on fine-tuning with whole-brain data. This is noted in the captions but we will revise this and make this distinction more explicit both in the main text and the caption.

Minor comments

Q1: We agree this would improve clarity and will include graphical illustrations of the requested components in the appendix for the camera-ready version.

Q2: We agree with the reviewer and will move the Training Protocol section to Appendix G in the camera-ready version.

Q3: Thank you for this suggestion. We would like to note that contrastive loss yielded better brain encoding performance (mean r = 0.163, SE = 0.0097), slightly outperforming Ridge (mean r = 0.1622) and MSE (mean r = 0.1612). Given that contrastive objectives are commonly used in prior brain decoding studies and offer strong alignment with brain representations (Défossez et al., 2023; Levy et al., 2025), we retained it as our main objective to ensure comparability with existing literature. We will revise the reference to Supplementary Table 1 and include this reasoning in the revised version for clarity.

Q4: Thanks for highlighting this; we indeed overlooked this. We’ll improve label visibility, enlarge ROI names, and reference Figures 3 and 4 earlier to ensure clarity.

Clarification on the Main Objective of the Paper

Thank you for your comment. Prior studies have fine-tuned speech-based language models with brain data and shown improvements in encoding performance (Moussa et al., 2025; Vattikonda et al., 2025). In contrast, our study focuses on text-based language models, which already exhibit strong semantic capabilities and have been shown to align well with brain activity. The key question we address is whether bilingual brains can serve as a useful supervisory signal to enhance the multilingual capabilities of language models. Specifically, we ask whether fine-tuning with a bilingual brain, which is known to encode a shared semantic space across languages (Chen et al., 2024), can benefit both monolingual and multilingual language models.

The comparison between fine-tuning with monolingual and bilingual brain data serves primarily as a control to assess whether the observed improvements in multilingual performance are attributable to brain data in general, or specifically to the bilingual brain.

Clarification on Methods

Thank you for the helpful feedback. We will include additional citations and elaborations in the revised version to explain the foundational concepts more clearly for a broader audience. To clarify a few terms: “brain encoding” refers to the prediction accuracy of voxel-wise timecourses, where each voxel in our data corresponds to a 2.24 × 2.24 × 4.1 mm³ volume of the brain. “TR” stands for repetition time, which is the acquisition time for each fMRI volume; in our case, the TR was 2.0045 seconds.

Clarification on design decisions

Response for Q1

For speech-based models, Vattikonda et al. (2025) showed that fine-tuning with different brain regions can yield differences, as regions may compete during fine-tuning and dominate the loss (see also Antonello et al., 2023). To avoid noise or voxels from non-language/semantic areas dominating the loss, we fine-tuned two complementary models for each participant: one using only brain responses from language regions and another using semantically-selective regions. The semantically-selective regions were chosen in a data-driven manner for each participant using encoding performance, while the language regions were defined using a group-level template (Fedorenko et al., 2010). These variants showed slightly better performance than the whole-brain model on some tasks, as reported in Appendix Section F.

Response for Q2

Thank you for the insightful question. Typically in fMRI studies, text stimuli are presented either in a controlled manner for example as a fixed number of words per repetition time (TR) (e.g., four words per TR, each lasting 0.5 seconds used in (Schwartz et al. 2019)) or as continuous input later downsampled during preprocessing. This limits flexibility, especially in settings where word timing is uneven. Performing stimulus-TR alignment as a preprocessing step (prior to fine-tuning) likely discards valuable temporal and contextual information.

In contrast, our pipeline introduces stimulus-TR alignment during fine-tuning, which has three key advantages:

• variable-length input (since word timing per repetition time (TR) can vary)
• more flexible and generalizable alignment applicable across modalities and tasks as we pass raw text as input
• dynamic stimulus-repetition time (TR) alignment as part of the learning process, not a static preprocessing step

In response to your suggestion, we implemented an alternative version of our pipeline that mirrors the preprocessing-based approach used in prior work; stimulus-TR alignment is performed prior to fine-tuning. Specifically, we constructed stimulus-TR pairs similar to Schwartz et al. (2019). We then fine-tuned the language model using this pre-aligned input and bilingual brain data from 1 participant. The mean encoding performance is 0.13 for the preprocessing-based method, compared to 0.15 for the vanilla model and 0.16 for our fine-tuned model. The table below presents downstream task performance (on the GLUE benchmark) for this preprocessing-based method (denoted as BERT-ft-en (Ablation)). This allows direct comparison with our proposed end-to-end brain-informed fine-tuning pipeline. We observe a substantial drop in performance for the preprocessing-based method (BERT-ft-en (Ablation)). This suggests that stimulus-TR alignment performed prior to fine-tuning procedure likely discards valuable temporal and contextual information. In contrast, our approach integrates brain alignment during fine-tuning, enabling robust downstream performance.

Clarification on baseline for Voxelwise encoding modeling

We agree that strong baselines are important for comparing the value of fine-tuning with bilingual brain data. Our current baselines are designed to test two specific hypotheses. First, to assess whether the fMRI signal alone, without meaningful temporal alignment to language, is sufficient. For this, we fine-tuned with temporally misaligned fMRI data. Second, to test if encoding improvements could stem from a different multilingual representation without brain data, we used mBERT embeddings as a baseline.

Clarification on Downstream tasks

Response for Q1

Thank you for this helpful question. Models were fine-tuned separately for each participant and language (note Methods 3.3). We will explicitly highlight this in the paper to ensure clarity. The numbers reported in Table 1 correspond to Participant 1. We have revised Table 1 to report the average performance across all six bilingual participants, along with standard deviations. Due to space constraints, please see the updated table 1 included in our response to reviewer hVbZ. As shown in the updated results, the overall trends remain consistent with those originally reported in the paper. We appreciate the suggestion since this change will help make the findings more interpretable and robust.

Response for Q2 and Q3

We agree with the concern regarding the small differences and sample size reported in Tables 1 and 2.

Regarding Table 1, we have revised Table 1 to report the average performance across all six bilingual participants, along with standard deviations (see comment above).

Regarding Table 2, we agree that the small sample size limited the strength of the conclusions. In response to your feedback, we identified two additional monolingual participants from another dataset (LeBel et. al, 2023) and have now included them in the analysis. Due to space constraints, please see the updated table 2 included in our response to reviewer hVbZ. With this expanded comparison group (3 monolingual participants), the results are consistent across participants: bilingual brain fine-tuning consistently outperforms the mBERT baseline across tasks and offers slightly stronger generalization than fine-tuning on individual monolingual participants. Bilingual brain fine-tuning outperforms monolingual fine-tuning on 7 out of 8 tasks on the CLUE benchmark, with equal performance on the remaining task (ChID). We thank the reviewer for this suggestion as we think this extended analysis better supports the contribution of our paper and makes our results more convincing.

[LeBel et al., 2023] A natural language fMRI dataset for voxelwise encoding models

Clarification on model choice

Please note that the supplementary material is indeed different from the appendix; it is separately downloadable. The supplementary material (referred to in the main text) includes comparison results for various model architectures: across mBERT, XLM-R, and XGLM. We apologize for the oversight and have now also included results for LLaMA-3.2-1B.

Due to space constraints, please see the results and comparison for mBERT, XLM-R, and XGLM (updated Supplementary Table 2) in our response to reviewer xVwE. This table shows that fine-tuning improves performance on most tasks: mBERT (7/9 tasks), XLM-R (9/9), and XGLM (6/9), with especially large gains for XLM-R (e.g., RTE: 54.15 → 62.71, MRPC: 69.61 → 77.94).

Due to space constraints, please see the results for LLaMA-3.2-1B in our response to reviewer 4NET. Bilingual brain fine-tuning outperforms the vanilla model on 2 out of 3 tasks. Please note that, due to computational constraints, LLaMA was fine-tuned using LoRA only at layer 9. As a result, the observed performance gains on downstream tasks are low compared to the other models.

Thank you for highlighting the points we had overlooked. Your feedback has helped us improve the clarity of the paper. We will incorporate all the tables and explanations provided above into the camera-ready version.