acceptodds
Under review as a conference paper at ICLR 2027

From Activation to Specificity: Automating Counterfactual Testing of Visual Representations in the Human Brain

Abstract

Identifying which brain regions represent a visual concept in the human brain is a central neuroscience challenge. Existing approaches have localized coarse functional regions (e.g., faces, places) through activation maximization, identifying regions that activate strongly for a target concept relative to others. Yet strong activation alone does not establish that a region represents the concept itself, as responses may instead be driven by correlated visual or semantic cues. We introduce BrainTRACE (Testing Representations through Counterfactual Evidence), an automated framework that combines generative and brain models to synthesize controlled stimuli and test candidate neural representations through targeted counterfactual-specificity testing. Given a query specifying a concept, our framework constructs targeted stimulus sets comprising concept images, counterfactual edits that remove the target concept while preserving other image content, and images with correlated distractors. It then uses an image-to-fMRI encoding model to predict brain responses and searches for representations that respond specifically to the target concept over correlated alternatives. BrainTRACE returns candidate representations supported by counterfactual-specificity evidence and proposes follow-up fMRI experiments to further test or extend these findings. Our approach recovers known functional localizations and identifies new candidate representations across dozens of concepts, supported by both predicted and measured fMRI data. Critically, we show that without counterfactual evaluation, a large fraction of localizations would be identified as false positives by our counterfactual-specificity testing, suggesting that activation alone is insufficient evidence of representation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Computational neuroscience, fMRI, visual concept representations, neural representation discovery, counterfactual evaluation, brain encoding models

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_DRHfOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses confounding in visual regions of the brain tied to specific concepts. Using NSD, which is full of the typical confounds of naturalistic images, lots of prior research has come out to try to map what the fMRI data says about how the brain represents concepts. Much of this research is tricked by confounds and is likely noise. This paper aims to tighten up the methodology and analysis by looking at context-preserving / manipulations that look for regions of the brain that would respond, e.g. equally to a concept and the same environment without that concept. This would mean, while older methods (activation maximisation) would incorrectly label an area as responsive to a concept, the proposed method would be able to detect this and falsify the claim. The proposed method is BrainTRACE. The paper largely confirms what's been said in other papers and I think slightly tries to oversell its contributions (perhaps in a defensible way and my own reading of certain sentences expected too much, e.g. that newly acquired fMRI data was used to confirm hypotheses). There should have been a Limitations section in here because the Conclusions section doesn't talk about any interesting useful findings and it's largely centred around limitations, not actually concluding much at all except restating the importance of the analysis (which I don't disagree with). Overall, while not maybe providing solid new insights, it does provide some extra scientific analysis to support regions that other papers have found, which we now know survive a strong counterfactual analysis.

Critical strengths
  • Tackles a real and important problem (I like papers that take a stance on attacking looser methodology seen elsewhere)
  • The motivation is sound and the literature is mostly well covered.
  • Methodological care against double dipping: voxels are selected on a training split and evaluated on held-out images.
  • I liked the conservative scoring on taking the most difficult / tricky potential edits and using those to derive scores
  • Good application of standard stats expected in the field, verifies my confidence a team actively working in the field did this
  • Sanity checks included to verify data quality were applied
Critical weaknesses
  • I had so many questions about the image-to-fMRi encoder and was shocked to see it not referenced in the main paper but banished to the appendices (I appredciate there are space concerns) but I felt it was important to talk more about that to better understand how a potentially confounded model, on same subjects and dataset, might also reveal confounding problems. I'm not 100% satisfied I have fully been able to convince myself there is no issue here with this potential circularity (this is not a criticism, just a concern I couldn't do away with after reading the paper).
  • The abstract implies the new approach is supported by new fMRI data, but no new fMRI data is collected, everything is based on the image-to-fMRI encoder predictions, so without that verification I also have a problem with the paper title (testing of visual representations in the human brain) -- yes, it's NSD, yes, the theories can be applied to the human brain but that's not what actually happens in this paper, so I feel there is a bit of overselling going on
  • I fundamentally didn't get any new insights about representations in the brain from this, though. We've all seen variations on these plots before, flatmaps of the same NSD areas and the associated pictures, large coarse-grained categories. The goal of the paper was to support strengthened conclusions (which was good) but the part of the paper saying about human hands, legs, feet, tools being somewhat 'fine-grained visual concepts' is not something I'd say I agree with.
  • Conclusion really was underselling the main aim of the paper and reminded the reader of all the caution that is warranted with an approach that relies on another model and no actual fMRI data to support the new conclusions (I'm guessing that's coming, and that would be a very interesting addition).
Questions for authors
  1. In the image-to-fMRI model, I'm a bit wary about the cross-over between the encoder model seeing images and fMRI data later used separately to derive your own conclusions about it. Your submission is quite weak in not always specifying the details on this (however, I appreciate the space concerns you are under). Can you tell me about how you effectively rationalised over this issue and what steps you took to make sure it didn't weaken your analysis in any way?

2.The failed semantic negatives, is this just a comment on the fact the pipeline failed in those instances or were they used in the analysis and then discovered post-hoc? I didn't quite get that bit.

  1. Are there plans to continue any experiments and collect fMRI data based on this?
3 comments

We thank the reviewer for their thoughtful feedback and the time devoted to reviewing our paper. We have made every effort to address each concern through additional experiments, clarifications, and revisions to the manuscript. The new experiments have been incorporated into the revised appendix and are discussed or referenced in the main paper where appropriate. All changes are highlighted in blue.


W1. I had so many questions about the image-to-fMRi encoder and was shocked to see it not referenced in the main paper but banished to the appendices (I appredciate there are space concerns) but I felt it was important to talk more about that to better understand how a potentially confounded model, on same subjects and dataset, might also reveal confounding problems. I'm not 100% satisfied I have fully been able to convince myself there is no issue here with this potential circularity (this is not a criticism, just a concern I couldn't do away with after reading the paper).

We agree that the image-to-fMRI encoder and its relationship to BrainTRACE’s evaluation deserve a clearer explanation. We have added a dedicated paragraph in Section 3 of the main paper and more detailed information in Appendix H.3. BrainTRACE uses the encoder of Beliy et al., trained jointly across all eight NSD subjects using only the BrainTRACE training splits, with shared NSD images across subjects excluded. Importantly, neither the validation images nor their measured responses are used for encoder training or representation discovery.

To further assess potential encoder-related biases, we evaluated BrainTRACE using an independently trained transformer encoder, obtaining highly consistent specificity maps (r = 0.868; Appendix D.1). We also added a new cross-dataset experiment using LAION-fMRI Subject 1, which differs from NSD in both its training-image distribution and measured fMRI responses. For each concept, BrainTRACE independently discovers a voxel set in each dataset and evaluates its specificity on held-out images. Despite the differences between brains, the specificity scores show strong agreement across 234 concepts (Spearman ρ = 0.80, computed across concepts rather than voxels), with 94.4% of strongly supported NSD concepts retaining positive specificity in LAION-fMRI (Appendix A).


W2. The abstract implies the new approach is supported by new fMRI data, but no new fMRI data is collected, everything is based on the image-to-fMRI encoder predictions, so without that verification I also have a problem with the paper title (testing of visual representations in the human brain) -- yes, it's NSD, yes, the theories can be applied to the human brain but that's not what actually happens in this paper, so I feel there is a bit of overselling going on

We would appreciate additional clarification to ensure that we fully understand your concern and make the necessary changes. BrainTRACE aims to explore and evaluate candidate visual representations in the brain using two complementary sources of evidence: predicted fMRI responses to generated positive images, semantic negatives, and counterfactual edits, and measured fMRI responses to positive and negative images retrieved from existing fMRI datasets. While counterfactual-specificity testing relies on encoder predictions, measured fMRI responses provide an independent evaluation of activation and semantic specificity. We acknowledge in the paper that existing measured data do not provide sufficient evidence for all concepts. Accordingly, BrainTRACE identifies these gaps and proposes targeted follow-up fMRI experiments to further test or extend its discoveries. To avoid suggesting that new fMRI data were collected, we have revised the abstract to replace "measured fMRI data" with "existing fMRI data". We hope this clarification addresses the reviewer’s concern and would be happy to make further revisions if needed, including changes to the abstract or rephrasing of the title.

W3. I fundamentally didn't get any new insights about representations in the brain from this, though. We've all seen variations on these plots before, flatmaps of the same NSD areas and the associated pictures, large coarse-grained categories. The goal of the paper was to support strengthened conclusions (which was good) but the part of the paper saying about human hands, legs, feet, tools being somewhat 'fine-grained visual concepts' is not something I'd say I agree with.

We agree that the main contribution of BrainTRACE is not necessarily the discovery of entirely new visual representations, but rather providing a systematic framework for evaluating and strengthening conclusions about their localization. One of our main findings is that many regions identified by activation-based methods fail counterfactual-specificity testing, demonstrating why stronger evidence is needed before interpreting high activation as concept representation. We use the term fine-grained to describe two complementary aspects of BrainTRACE: (i) More precise localization of concept-specific representations, rather than the broad regions of highly activated voxels typically identified by activation-based methods, as discussed in Appendix C.1. (ii) Better differentiation between closely related concepts, such as different human body parts, which often activate overlapping brain regions. By testing semantic alternatives and counterfactual edits, BrainTRACE helps distinguish representations that would otherwise be difficult to separate.


W4. Conclusion really was underselling the main aim of the paper and reminded the reader of all the caution that is warranted with an approach that relies on another model and no actual fMRI data to support the new conclusions (I'm guessing that's coming, and that would be a very interesting addition).

We agree and have strengthened the conclusion to better emphasize BrainTRACE’s main findings, particularly the importance of specificity testing for identifying more reliable candidate representations. We also clarify the dependence on the coverage of existing measured fMRI data and the use of predicted fMRI.


Q1. In the image-to-fMRI model, I'm a bit wary about the cross-over between the encoder model seeing images and fMRI data later used separately to derive your own conclusions about it. Your submission is quite weak in not always specifying the details on this (however, I appreciate the space concerns you are under). Can you tell me about how you effectively rationalised over this issue and what steps you took to make sure it didn't weaken your analysis in any way?

Please see our response to W1.


Q2. The failed semantic negatives, is this just a comment on the fact the pipeline failed in those instances or were they used in the analysis and then discovered post-hoc? I didn't quite get that bit.

The failed semantic negatives were included in the reported analysis and identified only post hoc, when we investigated cases where BrainTRACE still produced false positives. These examples were not excluded or corrected after inspection, but are discussed as a limitation of the current pipeline.


Q3. Are there plans to continue any experiments and collect fMRI data based on this?

We certainly plan to continue this work. We are already in contact with several groups, both within and outside our institution, sharing our method and findings and discussing potential fMRI experiments. These include targeted follow-up experiments to validate representations already explored by BrainTRACE, as well as discussions about which new concepts would be interesting to investigate using the pipeline.

Reviewer_DRHfOpenReview ↗

Thank you

Thank you for the care and time taken to address my concerns. I'll +1 my rating as recognition of the work put in to adapt the submission into what I believe is a better and more honest version of the paper.

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

BrainTRACE introduces a framework that test representations of encoding models through counterfactual evidence. That is done by testing whether voxels in Image-fMRI models not only respond strongly to visual concepts but also are specific to it. For each concept, an LLM writes prompts for three image sets: positive images, semantically related images, and images with minimal edits that remove the concept. Using these prompts, BrainTRACE generates the images using a image generative model, and verify it using vision language model. Results are checked on held-out generated images and on measured NSD data (4 subjects, 260 concepts). The aim to make these encoding (image-fMRI) models more concept-specific. Through BrainTRACE, authors find that:

  1. About 73% of concepts localized by activation-based ranking fail this test.
  2. BrainTRACE's activations are comparable, with higher specificity on generated and NSD data.
  3. Top voxels align with known face, body, place and word regions.
Critical strengths
  • A useful, reusable control. Existing neural localization approaches rely heavily on activation maximization, rendering them susceptible to false positives driven by correlated visual or semantic cues rather than the target concept itself.

  • Showing the 73.4% FPR for activation-based vs 23.0% FPR for specificity based approaches is very interesting result and shows that there a limitation that is worth addressing for future follow-up fMRI experiments.

Critical weaknesses
  • The manuscript do not cite MINE (Grosbard, Geva & Yovel, May 2026), which overlaps a lot with it: 1) The same dataset and subjects, 2) FLUX-based image generation and editing, 3) Counterfactual edits that add or remove features, 4) Fine-grained selectivity inside category-selective regions (FFA, EBA, PPA, word areas). The approaches differ in where hypotheses come from, MINE derives them bottom-up from an interpretable encoder, while, BrainTRACE starts from a query concept and adds semantic negatives, checks on measured data, and FDR tests. MINE describes its edit tests as “causal at the model level but correlational with respect to the brain.” Which contributions are not already covered by MINE? Where the two methods study the same subjects and regions (e.g., hand and body-part profiles in EBA/FBA), do their fine-grained maps agree? This previous work brings into question the novelty of this work. The authors could address my concerns if they can please elaborate on the distictions between their work and this previous work.

  • Counterfactual evidence is entirely on the image-fMRI model: Because NSD lacks controlled image edits, every counterfactual-specificity score reflects encoder predictions on FLUX.2-edited images rather than measured neural responses. Consequently, validation on recorded fMRI data is restricted to activation levels and semantic negatives (Table 1, Table T9). For Subject 1, while 57 of the 260 concepts pass FDR-corrected counterfactual testing, none can be verified against recorded brain activity. Because these statistical tests evaluate synthetic model outputs, their p-values reflect model statistical power rather than direct neural evidence.

  • The manuscript is missing a step that establish the accuracy on the 512 x 512 synthetic FLUX images or demonstrate that it can predict the subtle response difference between an image and its minimal edit. In the current state, it raises a out-of-distribution concern.

Questions for authors
  • Is there evidence that the encoder predicts responses well to generated images and to differences between an image and its edit?

  • For Table 2, which concepts were used per category, and how were they chosen?

  • Given that the manuscript repeatedly highlights BrainTRACE's capacity to design follow-up experiments, could you showcase a specific, concrete experiment generated by the system for a given concept and detail how its informativeness is quantified? Otherwise, could you please adjust the manuscript to frame this functionality as a theoretical proposal.

2 comments

We appreciate the reviewer’s careful evaluation and insightful comments. We provide responses to the concerns and questions raised below, supported by additional experiments. The new experiments have been incorporated into the revised appendix and are discussed or referenced in the main paper where appropriate.


W1. The manuscript do not cite MINE (Grosbard, Geva & Yovel, May 2026), which overlaps a lot with it [...]. The approaches differ in where hypotheses come from, MINE derives them bottom-up from an interpretable encoder, while, BrainTRACE starts from a query concept and adds semantic negatives, checks on measured data, and FDR tests. [...] Which contributions are not already covered by MINE? Where the two methods study the same subjects and regions (e.g., hand and body-part profiles in EBA/FBA), do their fine-grained maps agree? This previous work brings into question the novelty of this work. The authors could address my concerns if they can please elaborate on the distictions between their work and this previous work.

We thank the reviewer for bringing MINE to our attention. We would like to note that the two papers were developed concurrently, with their first preprints appearing in the same month. We therefore ask that MINE be considered concurrent rather than prior work. Nevertheless, we believe BrainTRACE makes several important and distinct contributions.

At a high level, MINE aims to interpret the features driving individual voxel predictions through mechanistic analysis of its encoder, whereas BrainTRACE aims to systematically evaluate whether independently specified visual concepts have reliable neural representations, quantify failures of activation-based localization, and identify the additional evidence needed to validate candidate representations.

More specifically: 1. Formulating and quantifying false-positive concept localization. While MINE is also motivated by limitations of activation-based interpretations, BrainTRACE explicitly formulates this problem and introduces a systematic quantitative framework for testing previously and newly discovered representations. 2. Investigating any researcher-defined concept. BrainTRACE tests arbitrary queried concepts independently of the encoder's interpretability mechanism, enabling investigation of otherwise overlooked concepts and rejection of unsupported localizations. 3. Systematically testing hard semantic negatives. Beyond counterfactual edits, BrainTRACE tests related and correlated concepts that may explain activation but are not necessarily ruled out by edits. 4. Validation using measured fMRI. BrainTRACE retrieves verified positives and hard negatives from measured fMRI datasets, providing independent neural evidence for concept specificity. 5. Guiding follow-up fMRI experiments. BrainTRACE characterizes and identifies gaps in measured data and proposes targeted stimuli for future experiments. 6. Extensive evaluation strengthening reliability. We demonstrate robustness across alternative brain encoders, language and image models, and fMRI datasets (Appendix D). We have referenced MINE in the revised manuscript.


W2. Counterfactual evidence is entirely on the image-fMRI model: Because NSD lacks controlled image edits, every counterfactual-specificity score reflects encoder predictions on FLUX.2-edited images rather than measured neural responses. Consequently, validation on recorded fMRI data is restricted to activation levels and semantic negatives (Table 1, Table T9). For Subject 1, while 57 of the 260 concepts pass FDR-corrected counterfactual testing, none can be verified against recorded brain activity. Because these statistical tests evaluate synthetic model outputs, their p-values reflect model statistical power rather than direct neural evidence.

A central motivation behind BrainTRACE is to extract more reliable neuroscientific insights from existing fMRI while guiding experiments. Since collecting new fMRI measurements is expensive and accessible to relatively few laboratories, much of computational neuroscience focuses on extracting additional knowledge from existing measurements, including through in silico predictions of fMRI responses to unseen stimuli. BrainTRACE goes further by systematically testing concept specificity using generated counterfactual edits and hard semantic negatives, alongside validation on positive and negative images retrieved from measured fMRI responses. We acknowledge that direct fMRI measurements of generated edits are needed for further validation, and an important goal of BrainTRACE is to identify which concepts and stimuli should be prioritized for such experiments. Following the reviewer's suggestions, we have now added multiple experiments to further assess the reliability of predicted fMRI responses to generated edits (see W3), as well as a follow-up experimental design pipeline with illustrative examples (see Q3).

W3. The manuscript is missing a step that establish the accuracy on the 512 x 512 synthetic FLUX images or demonstrate that it can predict the subtle response difference between an image and its minimal edit. In the current state, it raises a out-of-distribution concern.

We thank the reviewer for raising this concern, which motivated additional experiments examining the encoder's generalization and its sensitivity to changes introduced by generated counterfactual edits.

First, we show that predicted fMRI response differences capture intervention-specific changes. Using BrainTRACE's original human-face counterfactuals, we trained a linear decoder to distinguish five types of edits (e.g., replacing, covering, or removing a face) from predicted original-minus-edit responses. On 97 edits from entirely unseen source images, the decoder achieved 79.7% balanced accuracy (20% chance). We further tested a broader set of controlled FLUX.2 edits, evaluating whether predicted response differences distinguish which object replaces another. Across seven six-class replacement tasks (e.g., fruits, animals, and utensils), decoding on unseen source scenes achieved 80.8–100.0% balanced accuracy (16.7% chance), consistently outperforming a coarse pixel-difference baseline (Appendix E.3; Table T10, Fig. S5).

Second, we demonstrate that the encoder generalizes to non-natural out-of-distribution (OOD) images with measured fMRI validation. To evaluate generalization beyond the original NSD dataset, we transferred the encoder to LAION-fMRI, whose held-out test set includes OOD, non-natural stimuli (e.g., visual illusions, Gabor patches, and geometric shapes). The encoder achieved median voxelwise correlations of 0.500 in-distribution and 0.441 OOD, demonstrating substantial predictive performance on unseen stimulus types (Appendix E.1).

Third, we show that the encoder can distinguish visually and semantically similar images. For each of 1,000 held-out NSD images, we used its measured fMRI response to retrieve the correct image by ranking the encoder-predicted responses of all 1,000 candidates. The correct image was ranked first in 94.0% of cases (97.5% Top-5), showing to capture fine-grained visual distinctions between highly similar alternatives (Fig. S4; Appendix E.2).

Together, these experiments provide stronger evidence for the reliability of predicted counterfactual effects and their use in guiding targeted fMRI experiments for direct neural validation (see Q3).


Q1.Is there evidence that the encoder predicts responses well to generated images and to differences between an image and its edit?

Please see our response to W3.


Q2. For Table 2, which concepts were used per category, and how were they chosen?

Table 2 includes 12 concepts across four fLoc categories: bodies (body part, human hand, human leg), faces (human face, eyes, expressive face), places (bedroom, kitchen, bathroom), and words (text, printed word, handwritten text), selected to cover both broad and fine-grained visual distinctions within established functional categories. We will include the full list in the revised manuscript.


Q3. Given that the manuscript repeatedly highlights BrainTRACE's capacity to design follow-up experiments, could you showcase a specific, concrete experiment generated by the system for a given concept and detail how its informativeness is quantified? Otherwise, could you please adjust the manuscript to frame this functionality as a theoretical proposal.

We have clarified BrainTRACE’s follow-up experiment design procedure, detailing how it assesses the available measured evidence, identifies missing stimuli, and proposes informative panels for future fMRI measurements. We additionally added two concrete examples, for Body in Motion and People Cuddling in Appendix F.4 (Fig. S7 and Table T14). BrainTRACE evaluates whether existing measured fMRI data provide sufficient evidence for a discovered representation, considering both statistical significance and stimulus coverage. When evidence is insufficient, it proposes a panel of positives, semantic negatives, and counterfactual edits. Using predicted fMRI responses, it selects the smallest tested panel for which activation and both specificity tests reach nominal significance (p < 0.05).

Reviewer_s2LmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a BrainTRACE framework for discovering visual concept representations in fMRI data using counterfactual-specificity testing. Instead of ranking voxels only by activation to positive images, it generates three stimulus types—positives, semantic negatives, and counterfactual edits—and selects voxels that respond strongly to the target concept but less strongly to correlated alternatives. It validates candidates on held-out predicted and measured fMRI data and proposes follow-up experiments when measured coverage is insufficient.

Critical strengths
  1. This paper proposes an interesting framework for activation-based concept localization with a three-stage design of targeted stimulus generation, concept-selective representation search, and validation/follow-up experiment design.
  2. Discovery uses predicted responses to targeted, generated stimuli, but evaluation is also performed on held-out measured NSD images where available. This is a meaningful validation strategy.
  3. The paper is clearly written and easy to follow.
Critical weaknesses
  1. The study relies heavily on the capability of the encoding model. However, the generalization of the encoding model to unseen images or out-of-distribution concepts is not validated. While the paper reports high correlations between predicted and measured responses, these results do not fully establish the reliability of predicted neural responses under counterfactual interventions.
  2. The paper assumes that voxels exhibiting higher activation for positive stimuli than for counterfactual negatives provide stronger evidence of concept-specific representations. However, this distinction is not necessarily sufficient to establish that a voxel represents the target semantic concept.
  3. The paper claims to discover fine-grained representations, such as different body parts and types of text. However, the evidence primarily consists of qualitative brain maps and alignment with established broad functional regions.
Questions for authors
  1. How reliable are the image-to-fMRI encoding models specifically on generated counterfactual stimuli, rather than ordinary natural images?
  2. Can the authors provide stronger controlled experiments in which semantic identity changes while low-level and intermediate-level visual properties remain closely matched?
  3. Do fine-grained concept maps exhibit statistically significant differences beyond the known broad category-selective regions? And how to quantitatively evaluate?
4 comments

We thank the reviewer for their thoughtful feedback and the time devoted to reviewing our paper. We have made every effort to address each concern through additional experiments and clarifications. The new experiments have been incorporated into the revised appendix and are discussed or referenced in the main paper where appropriate. All changes are highlighted in blue.


W1/Q1: W1. The study relies heavily on the capability of the encoding model. However, the generalization of the encoding model to unseen images or out-of-distribution concepts is not validated. While the paper reports high correlations between predicted and measured responses, these results do not fully establish the reliability of predicted neural responses under counterfactual interventions. Q1. How reliable are the image-to-fMRI encoding models specifically on generated counterfactual stimuli, rather than ordinary natural images?

We thank the reviewer for raising this important concern, which motivated several additional experiments. These experiments complement the encoder and generator robustness analyses already reported in Appendix D1 and further investigate the reliability of encoder predictions under counterfactual interventions.

(i) Generalization to out-of-distribution stimuli. To evaluate generalization beyond the original NSD dataset, we transferred the encoder to LAION-fMRI, whose held-out test set includes OOD, non-natural stimuli (e.g., visual illusions, Gabor patches, and geometric shapes). The encoder achieved median voxelwise correlations of 0.500 on in-distribution images and 0.441 on OOD images (Appendix E.1), demonstrating substantial predictive performance on unseen stimulus types.

(ii) Discriminating similar images. We tested whether the encoder can distinguish individual images even when they are highly similar visually and semantically. For each of 1,000 held-out NSD images, we used its measured fMRI response to rank the encoder-predicted responses to all 1,000 candidate images. The correct image was ranked first in 94.0% of cases (97.5% Top-5), capturing fine-grained visual distinctions between highly similar alternatives (Fig. S4, Appendix E.2).

(iii) Sensitivity to BrainTRACE counterfactual edits. We tested whether predicted fMRI response differences between original images and their counterfactual edits capture the specific changes introduced by editing. Using BrainTRACE's original human-face counterfactuals, we trained a linear decoder to distinguish five edit types (e.g., replacing a face with an animal head, covering it, or removing it) based on predicted fMRI response differences. On 97 edits from unseen source images, the decoder achieved 79.7% balanced accuracy (20% chance), indicating that encoder predictions capture intervention-specific information (Appendix E.3).

(iv) Sensitivity to fine-grained semantic changes. We extended this analysis to a broader set of controlled FLUX.2 edits, testing whether predicted fMRI response differences distinguish between object replacements. Across seven six-class replacement tasks (e.g., fruits, animals, and utensils), linear decoders trained on predicted edited-minus-original responses achieved 80.8–100.0% balanced accuracy (16.7% chance) on unseen source scenes (Appendix E.3; Table T10, Fig. S5). Overall, these results strengthen confidence in encoder-based counterfactual testing, while BrainTRACE guides future fMRI experiments for direct validation.

W2/Q2: W2.The paper assumes that voxels exhibiting higher activation for positive stimuli than for counterfactual negatives provide stronger evidence of concept-specific representations. However, this distinction is not necessarily sufficient to establish that a voxel represents the target semantic concept. Q2. Can the authors provide stronger controlled experiments in which semantic identity changes while low-level and intermediate-level visual properties remain closely matched?

Distinguishing responses driven by a target concept from those driven by correlated visual properties is a fundamental challenge in visual neuroscience, and a central motivation for BrainTRACE. Classical neuroscience studies have used controlled stimulus comparisons, but typically examine a limited number of concepts through a small set of manually designed contrasts. BrainTRACE extends this approach systematically, generating diverse counterfactual edits that remove the target concept while preserving low- and intermediate-level visual properties as closely as possible, alongside semantic negatives targeting correlated concepts. Moreover, across 67,517 original–edit pairs, we show that nine visual properties (e.g., luminance, contrast, color, edges, and spatial-frequency content) explain only 6.2% of the variance in specificity scores, which remain highly consistent after residualization (r = 0.97; Appendix C.5). We will further clarify that BrainTRACE provides stronger evidence of concept specificity than previous methods, while acknowledging that we cannot exclude every possible confound.


W3/Q3: W3. The paper claims to discover fine-grained representations, such as different body parts and types of text. However, the evidence primarily consists of qualitative brain maps and alignment with established broad functional regions. Q3. Do fine-grained concept maps exhibit statistically significant differences beyond the known broad category-selective regions? And how to quantitatively evaluate?

A central goal of BrainTRACE is to distinguish representations of a target concept from those driven by related concepts, using both hard semantic negatives and counterfactual edits. To further validate this ability at a fine-grained level, we conducted two additional quantitative experiments.

(i) Discriminating related concepts within broad functional ROIs. For each concept pair (e.g., human hands vs. human legs), we independently selected the top 100 voxels for each concept within the same ROI using the training split. On held-out images, we then tested whether the voxels selected for each concept responded more strongly to that concept than to its paired alternative, averaging the two directional contrasts. All six concept pairs showed significant discrimination on generated held-out images (BH-FDR q ≤ .0001). Importantly, all four comparisons with sufficient independently measured fMRI data were also significant (q ≤ .0012; Appendix F.2). (ii) Statistical significance of individual concept representations. On held out data, all six concepts showed significant semantic-specificity effects on measured fMRI (q ≤ .0033), and five showed significant counterfactual-specificity effects on predicted fMRI (q ≤ .049). Together, these results provide quantitative evidence for fine-grained functional distinctions within broad category-selective regions, beyond qualitative brain maps or alignment with known ROIs. We have incorporated these quantitative evaluations and clarified the findings in the revised manuscript (Appendix F.2).

Thank you for the rebuttal and experimental results. I still believe that the encoder generalization and the OOD issue of generated images have a great impact on the analysis experiments. On these points, the authors only partially addressed my concerns. I am inclined to keep my original score.

We thank the reviewer for their follow-up and for the time and effort devoted to reviewing our paper. We understand the reviewer's concerns regarding encoder generalization to generated images and have conducted an additional experiment to further address them, showing that the generated images are relatively well covered by the encoder training set image distribution (see below). Together with our previous and newly added experiments, we believe this substantially mitigates concerns about encoder generalization and OOD generated images, although we acknowledge that, as with other in-silico approaches, these concerns cannot be fully resolved without additional fMRI experiments.

We would also like to ask the reviewer to consider BrainTRACE's contributions independently of complete validation of the in-silico counterfactual results. As the reviewer noted in their strengths, our three-stage framework and validation using held-out measured fMRI provide meaningful contributions. We believe our problem formulation and retrieval-based specificity evaluation using measured fMRI positives and negatives already provide meaningful contributions independently of predictions for generated images. Together with in-silico counterfactual testing and guidance for prioritizing future fMRI experiments, we believe BrainTRACE provides a valuable framework and hope the reviewer will consider these contributions in their overall assessment.


New Experiment. Our previous experiments demonstrated that the brain encoder generalizes well to both held-out natural images and OOD LAION-fMRI stimuli. To further assess whether this generalization supports its use on BrainTRACE's generated images, we compared their visual distribution with the images used to train and evaluate our encoders. Specifically, we measured nearest-neighbor distances to a combined NSD and LAION-fMRI image pool, using DINOv2 (the same visual backbone used by our encoder) and CLIP features. We defined a distance threshold based on the 95th percentile of held-out natural images and measured the fraction of generated images exceeding it. In DINOv2, only 10.4% of positives, 12.8% of semantic negatives, and 10.1% of counterfactual edits exceeded this threshold, compared with 36.4% of LAION-fMRI OOD images. In CLIP, the corresponding rates were 6.5%, 4.3%, and 6.2%. Thus, our generated stimuli are considerably closer to the reference image distribution than the OOD stimuli on which the encoder already demonstrated strong predictive performance. This provides additional evidence that the generated images fall within a visual distribution over which our brain encoders have demonstrated meaningful generalization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.