acceptodds
Under review as a conference paper at ICLR 2027

When Model Complexity Matters in Cross-Subject fMRI Prediction

Abstract

Increasingly complex neural models predict blood-oxygen-level-dependent (BOLD) responses to naturalistic stimuli, but the predictive value of this added complexity remains unclear. We ask when pretrained stimulus representations help and how much downstream complexity is needed, with and without recent measured BOLD history. We compare linear diagnostic references and nonlinear predictors across HCP movie fMRI (154 subjects), BOLD Moments (10 subjects), and Algonauts 2025 (4 subjects). Under a shared offline protocol, models are fitted on training subjects and evaluated on held-out subjects viewing largely familiar stimuli, without refitting. Pretrained visual features generally improve prediction over raw pixels; for linear autoregression, their advantage grows as measured-history refreshes become less frequent. Measured BOLD history provides substantial gains that generally diminish with longer refresh intervals. Nonlinear stimulus-only predictors can improve on linear readouts, whereas linear autoregression provides a strong history-conditioned baseline. Controls examine context length, regularization, fitting procedures, and model capacity. Regional response consistency across subjects is associated with held-out prediction accuracy, while less consistent regions tend to gain more from measured history. In paired simulations with fixed linear neural dynamics, slower hemodynamic responses generally sustain high temporal prediction correlations over longer refresh intervals, illustrating how the observation process can contribute to predictive persistence. Together, these findings support evaluating stimulus representations and downstream complexity separately, using history-conditioned linear prediction as a diagnostic reference under matched inputs and measurement availability.

Primary area:
applications to neuroscience & cognitive science
Keywords:
fMRI encoding models, cross-subject generalization, naturalistic neuroimaging, multimodal pretrained representations, computational neuroscience

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_pp2XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates when model complexity benefits cross-subject fMRI prediction. Across three datasets, the authors show that pretrained stimulus representations improve prediction accuracy, while simple linear autoregressive models often outperform complex nonlinear models when recent BOLD history is available. These findings highlight the importance of strong linear baselines and suggest that increased model complexity does not necessarily lead to better performance.

Critical strengths

• The paper addresses an important question regarding when increased model complexity is beneficial for cross-subject fMRI prediction.

• The regional analyses and controlled hemodynamic simulations provide useful neuroscientific insights.

Critical weaknesses

• The performance results were given in Figures and graphs without exact numbers, which may limit follow-up studies for fair comparison.

• The number of samples is limited; BOLD Moments has 10 and Algonauts has only 4 samples. Small subject cohorts and limited statistical uncertainty analyses weaken the generalizability of some conclusions.Moreover, the substantial overlap of stimuli between training and test subjects limits conclusions about generalization to unseen stimuli.

• Several nonlinear models are substantially adapted or reimplemented, raising concerns about whether the comparisons fairly reflect their original capabilities.

• The paper reads more like a technical report than a research paper, as it primarily focuses on benchmarking existing models and analyzing their performance under different settings. While the extensive experiments provide useful insights, the methodological novelty and conceptual contributions are somewhat limited.

Questions for authors

• Could the authors evaluate the original pretrained models or provide stronger evidence that the benchmark adaptations preserve their intended capabilities?

• How well do the findings generalize to entirely unseen stimuli? Could stimulus overlap favor simpler models that exploit stimulus-specific temporal patterns?

• How much of RR-AR's performance can be explained by temporal correlation alone, without stimulus information?

• Could the authors extend their simulations to nonlinear neural dynamics to determine whether the advantage of linear prediction persists?

5 comments
AuthorsOpenReview ↗

Response to numerical evidence and generalization

General Question 1 — Complete numerical results (W1)

Summary. Where can the exact numerical results underlying the figures be found?

Response.

We have added the complete numerical results corresponding to the reported figures to Appendix K (Numerical Results) of the revised manuscript.

General Question 2 — Statistical evidence and generalization (W2 / Q2)

Summary. Do the sample sizes and statistical analyses support the conclusions, and do the findings generalize to unseen stimuli?

Response.

The independent units for statistical inference are held-out subjects: 32 HCP, 10 BOLD Moments, and four Algonauts participants. Prediction time points describe evaluation size but do not increase the independent subject count (Appendix A.8). HCP provides the larger cross-subject evaluation alongside the two smaller cohorts.

Appendix I now reports subject-level paired model comparisons with pointwise 95% confidence intervals and Holm-adjusted p-values, alongside analyses of the paper's twelve main conclusions. For example, in the shared-input benchmark, RR-AR outperforms MLP and GRU at H=1,2,4,8H=1,2,4,8 in all three datasets after correction.

Generalization to unseen stimuli. We added a stimulus-disjoint HCP evaluation: models are trained on Movie 1/2, selected on Movie 3, and tested on Movie 4, with repeated clips excluded across partitions. We test both previously seen subjects and entirely held-out subjects. The comparisons include RR-AR, history-only ridge, RR-stim, MLP, GRU, BrainVista, TRIBE v2, and both shared-input and official-code MIRAGE configurations.

History-conditioned models use preceding available BOLD observations, whereas stimulus-only models receive no measured BOLD. Representative mean correlations for held-out subjects viewing unseen stimuli are shown below. Stimulus-only scores are independent of HH.

HCP: unseen subjects and unseen stimuli

Hist. = history-only ridge; BV = BrainVista; S/O = shared-input/official-code. BV and TRIBE v2 use shared inputs.

HRR-ARHist.BVMLPGRURR-stimTRIBE v2MIRAGE SMIRAGE O
1.621.680.621.386.210.128.121.134.044
2.569.622.555.378.205.128.121.134.044
4.479.515.440.316.176.128.121.134.044
8.378.387.318.249.141.128.121.134.044
16.296.283.227.201.115.128.121.134.044
32.242.208.160.168.102.128.121.134.044
96.197.144.109.145.092.128.121.134.044
Never.172.098.071.135.087.128.121.134.044

At short HH, recent BOLD history supports high accuracy: history-only ridge matches or exceeds full-input RR-AR, while RR-AR matches or outperforms BrainVista and outperforms MLP and GRU. At long HH, history-only accuracy declines and adding stimulus improves prediction; RR-AR remains stronger than BrainVista, MLP, and GRU. Thus, the main patterns—strong linear prediction with recent history and increasing benefit from stimulus as history becomes distant—persist on unseen stimuli in both subject cohorts. Appendix J of the revised manuscript provides the complete curves (Figure 19), numerical results (Tables 22–23), familiar-stimulus comparison, and implementation and statistical details.


AuthorsOpenReview ↗

Response to official implementations

General Question 3 — Official model implementations (W3 / Q1)

Summary. Do the benchmark adaptations fairly reflect the original models, and what happens when official implementations are evaluated?

Response.

(1) Why adapt the models for this benchmark? We appreciate the reviewer's concern about the influence of implementation choices. Our original benchmark was designed to compare downstream predictors under common stimulus inputs and matched BOLD-history availability. For MIRAGE, we therefore retained the released downstream network while replacing its Qwen features with the common features used by the other predictors. For BrainVista, whose official code was unavailable when we implemented the benchmark, we implemented the model from the published description and integrated it into the same evaluation framework.

(2) Additional runs using the authors' released implementations. For MIRAGE, we restore native Qwen3-Omni representations and released model components, train the downstream predictor from scratch on training subjects, and average their readout heads for unseen subjects. For BrainVista, we replace our paper-based implementation with the released model components and official training procedure. Both evaluations retain our cross-subject protocol.

The additional BrainVista results cover all three datasets, and MIRAGE covers HCP and BOLD Moments. MIRAGE on Algonauts would require more than two weeks of feature extraction on our available hardware, exceeding the revision window.

(3) How do the official-code models compare with RR-AR? The tables compare manuscript (M) and official-code (O) results against RR-AR. Rows show H (TRs); entries are mean (subject SD). MIRAGE is stimulus-only, so its score repeats across H; a dash indicates an unavailable result.

HCP

HRR-ARBrainVista MBrainVista OMIRAGE MMIRAGE O
1.685(.032).618(.035).551(.028).267(.049).047(.013)
2.632(.031).563(.034).506(.029).267(.049).047(.013)
4.541(.033).472(.034).424(.031).267(.049).047(.013)
8.444(.033).377(.030).332(.032).267(.049).047(.013)
16.377(.034).310(.025).264(.030).267(.049).047(.013)
32.336(.037).270(.024).224(.031).267(.049).047(.013)
96.304(.045).241(.031).188(.032).267(.049).047(.013)
Never.294(.049).217(.031).178(.032).267(.049).047(.013)

BOLD Moments

HRR-ARBrainVista MBrainVista OMIRAGE MMIRAGE O
1.408(.039).305(.043).320(.028).090(.008).104(.012)
2.329(.032).227(.036).244(.024).090(.008).104(.012)
4.254(.021).156(.025).173(.016).090(.008).104(.012)
8.193(.015).109(.020).127(.011).090(.008).104(.012)
16.147(.010).082(.014).097(.007).090(.008).104(.012)
32.110(.008).058(.010).079(.008).090(.008).104(.012)
96.073(.008).029(.008).064(.008).090(.008).104(.012)
Never.054(.009).002(.001).058(.008).090(.008).104(.012)

Algonauts

HRR-ARBrainVista MBrainVista OMIRAGE MMIRAGE O
1.514(.033).438(.028).424(.030).240(.011)—
2.449(.024).356(.015).353(.021).240(.011)—
4.368(.016).259(.008).265(.012).240(.011)—
8.308(.013).193(.007).201(.010).240(.011)—
16.268(.011).152(.008).161(.011).240(.011)—
32.241(.011).125(.011).140(.014).240(.011)—
96.222(.011).096(.016).127(.018).240(.011)—
Never.212(.012).022(.008).136(.020).240(.011)—

Comparison with RR-AR. The official-code reruns retain RR-AR's advantage over BrainVista on HCP and Algonauts. On BOLD Moments, RR-AR performs better at short intervals, while the two models become close at long intervals and nearly match without refresh (0.054 versus 0.058). Official-code MIRAGE remains well below RR-AR on HCP; on BOLD Moments, it is close to RR-AR at H=32H=32 (0.104 versus 0.110) and performs better at H=96H=96 and without refresh. Overall, RR-AR remains a strong reference when recent measured history is available, while the benefit of nonlinear stimulus-based prediction becomes more apparent as that history grows distant.

An HCP audit found no scoring or duplicated-offset error. Head averaging reduced accuracy in a training diagnostic, but scores were low with both readouts; it is therefore only a possible contributor to the low score (Appendix K.4).

Implementation and evaluation details, aggregation procedures, and complete results are provided in Appendices B and K.4.


AuthorsOpenReview ↗

Response to ML contribution and the role of history

General Question 4 — Contribution to machine learning (W4)

Summary. What empirical and conceptual contribution does the study make beyond benchmarking existing models?

Response. Thank you for pointing this out. ICLR’s reviewer guidelines explicitly recognize new, relevant, and impactful knowledge, including empirical findings, as a contribution. Methodological novelty is therefore not the only form of contribution recognized by ICLR.

Our study contributes new knowledge through three empirical insights:

  1. Representation quality and predictor complexity provide different benefits. Pretrained stimulus representations generally improve prediction, while increasing downstream complexity does not consistently improve performance under matched inputs.
  2. The value of model complexity depends on available history. Recent BOLD supports strong linear prediction, while stimulus input becomes more valuable as that history grows distant. History-only controls and the information analysis in Appendix H characterize these contributions.
  3. The observation process shapes predictability. Paired simulations show that slower hemodynamic responses can prolong history-based predictability, including under nonlinear neural dynamics (Appendices F–G).

This contribution is consistent with empirical representation studies published at ICLR:

Our study complements these representation studies by examining when additional predictor complexity improves BOLD prediction under matched stimulus representations and history availability.

General Question 5 — Predictive contribution of BOLD history (Q3)

Summary. How much of RR-AR's prediction accuracy is supported by BOLD history without stimulus input?

Response.

We refit RR-AR with stimulus inputs set to zero during training, validation, and testing, keeping the same BOLD history, subjects, and evaluation targets. This history-only control measures how much predictive accuracy is supported by past BOLD alone. Figure 12 and Appendix K (Numerical Results) of the revised manuscript provide the complete curves and numerical results. The tables below reproduce mean temporal correlations at representative intervals.

HCP

Inputs / H1248163296Never
History + stimulus0.6850.6320.5410.4440.3770.3360.3040.294
History only0.6820.6240.5170.3920.2940.2220.1490.120

BOLD Moments

Inputs / H1248163296Never
History + stimulus0.4080.3290.2540.1930.1470.1100.0730.054
History only0.4190.3270.2400.1760.1320.0950.0520.027

Algonauts

Inputs / H1248163296Never
History + stimulus0.5140.4490.3680.3080.2680.2410.2220.212
History only0.5070.4260.3190.2340.1720.1220.0730.035

At the shortest intervals, history-only RR-AR performs close to the full-input model across all three datasets. As HH increases, history-only accuracy declines, while the full-input model retains higher accuracy. The growing benefit of adding stimulus is especially apparent in HCP and Algonauts; the difference is smaller in BOLD Moments.

Thus, recent BOLD history accounts for much of RR-AR's short-interval predictive accuracy, while stimulus input becomes more valuable as that history grows distant.


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

The paper addresses the problem of fMRI encoding from video stimuli, i.e., predicting the fMRI brain response from video clips. Specifically, the paper examines prediction of the brain response from previous fMRI samples and the current stimulus using linear autoregression. It compares the performance against previously published models that predict from the stimulus only, and against preprints that predict from both the stimulus and previous fMRI samples.

Critical strengths

The paper examines the problem of encoding conditioned on both the stimulus and fMRI history, a problem that is less examined today, and performs a verry thorough analysis across multiple models and multiple datasets. It establishes a clear, strong, and simple reference that future complex models can be compared against.

Critical weaknesses
  • For an ML venue, the paper does not provide much in terms of advancing ML, as it mainly uses classical ML methods. A neuroscience journal may be a more appropriate venue.

  • I see this as more of a review paper on video encoding models conditioned on fMRI history. The methods covered (BrainVista and NeuroWorld) are preprints and do not really require a review at this point in time, and not testing a simple linear regression approach baseline is more a shortcoming of those works.

  • The overlap in stimuli between train and test is a methodological flaw. Since the test subjects watch the same stimuli seen in training, the models do not need to learn a mapping from the stimulus to the brain response; they can rely on memorizing the training subjects’ responses to these specific stimuli. In this setting a stimulus encoder is not really needed: one can simply take the mean response over the training subjects, and a possibly even better approach would be a nearest-neighbor (k-NN) search on the fMRI history, producing the prediction as a weighted average of the training subjects’ responses. Basically, any model with high enough capacity will work well. As a result, the evaluation does not test generalization to new stimuli, which is the main purpose of an encoding model, and the conclusions about which models and features matter may not hold in that setting. At minimum, these simple baselines should be reported.

  • Constraining the nonlinear models to the parameter count of the linear model does not make sense from a modern ML perspective and goes against what we currently know.

  • There is no need to invent new terms when established ones exist, e.g., “refresh interval” instead of “prediction horizon”. This only makes the paper harder to read.

Questions for authors

No additional questions.

6 comments
AuthorsOpenReview ↗

Response to original contribution and ML relevance

General Question 1 — Original contribution and ML relevance (W1 / W2)

Summary. What is the original research contribution, and why is it relevant to a machine-learning venue?

Response. Thank you for pointing this out. ICLR’s reviewer guidelines explicitly recognize new, relevant, and impactful knowledge, including empirical findings, as a contribution. Methodological novelty is therefore not the only form of contribution recognized by ICLR.

Our study contributes new knowledge through three empirical insights:

  1. Representation quality and predictor complexity provide different benefits. Pretrained stimulus representations generally improve prediction, while increasing downstream complexity does not consistently improve performance under matched inputs.
  2. The value of model complexity depends on available history. Recent BOLD supports strong linear prediction, while stimulus input becomes more valuable as that history grows distant. History-only controls and the information analysis in Appendix H characterize these contributions.
  3. The observation process shapes predictability. Paired simulations show that slower hemodynamic responses can prolong history-based predictability, including under nonlinear neural dynamics (Appendices F–G).

This is an original controlled experimental study across three datasets. Classical linear predictors serve as diagnostic references; BrainVista and NeuroWorld are comparison models within a broader assessment of model design.

This contribution is consistent with empirical representation studies published at ICLR:

Our study complements these representation studies by examining when additional predictor complexity improves BOLD prediction under matched stimulus representations and history availability.


AuthorsOpenReview ↗

Response to baselines and generalization

General Question 2 — Simple baselines and unseen stimuli (W3)

Summary. Can simple baselines explain the results, and do the findings generalize to unseen stimuli?

Response.

We evaluated both requested baselines—training-subject mean and history k-NN—on HCP, using the same held-out subjects and prediction targets as RR-AR. These controls test whether averaging or retrieving training responses reproduces its accuracy in the familiar-stimulus benchmark. We also added a stimulus-disjoint HCP evaluation to test generalization to unseen stimuli.

(1) Does a training-subject mean explain the prediction accuracy? This baseline averages training subjects' BOLD responses to the same stimulus version, clip, and time point (19 or 88 subjects, depending on version). It receives no test-subject BOLD and is independent of HH.

We compare it with RR-AR on the same 576 records from 32 held-out subjects. Values are mean ±\pm subject SD; construction and aggregation details are in Appendix K.6.

Predictor / H1832128Never
Training-subject mean0.26075 ± 0.047990.26075 ± 0.047990.26075 ± 0.047990.26075 ± 0.047990.26075 ± 0.04799
RR-AR0.68513 ± 0.031530.44415 ± 0.032710.33554 ± 0.037270.30044 ± 0.046480.29356 ± 0.04889

The mean-response baseline captures familiar-stimulus predictability but remains below RR-AR at every reported interval, including no refresh. Averaging training subjects' responses therefore does not reproduce RR-AR's accuracy.

(2) Can history-based retrieval reproduce the result? The training-only library contains 343,898 eight-frame windows from 107 subjects. Exact retrieval uses the full standardized history; validation selects Euclidean distance, k=20k=20, and inverse-distance weighting. Predictions update the history between measured observations. We evaluate all 14 refresh settings on the same 576 test records; Appendix K.6 gives the full procedure.

Predictor / H1832128Never
Exact history k-NN0.38384 ± 0.049000.17109 ± 0.037510.07090 ± 0.017600.03487 ± 0.010410.03110 ± 0.01043
RR-AR0.68513 ± 0.031530.44415 ± 0.032710.33554 ± 0.037270.30044 ± 0.046480.29356 ± 0.04889

Values are mean ±\pm sample SD across subjects, with the same reduction as above. This training-only, history-based retrieval baseline is below RR-AR at the reported settings. Together, the two proposed baselines capture part of familiar-stimulus predictability, while RR-AR remains more accurate on the same held-out subjects and targets.

(3) Generalization to unseen stimuli. We added a stimulus-disjoint HCP evaluation: models are trained on Movie 1/2, selected on Movie 3, and tested on Movie 4, with repeated clips excluded across partitions. We test both previously seen subjects and entirely held-out subjects. The comparisons include RR-AR, history-only ridge, RR-stim, MLP, GRU, BrainVista, TRIBE v2, and both shared-input and official-code MIRAGE configurations.

History-conditioned models use preceding available BOLD observations, whereas stimulus-only models receive no measured BOLD. Representative mean correlations for held-out subjects viewing unseen stimuli are shown below. Stimulus-only scores are independent of HH.

HCP: unseen subjects and unseen stimuli

Hist. = history-only ridge; BV = BrainVista; S/O = shared-input/official-code. BV and TRIBE v2 use shared inputs.

HRR-ARHist.BVMLPGRURR-stimTRIBE v2MIRAGE SMIRAGE O
1.621.680.621.386.210.128.121.134.044
2.569.622.555.378.205.128.121.134.044
4.479.515.440.316.176.128.121.134.044
8.378.387.318.249.141.128.121.134.044
16.296.283.227.201.115.128.121.134.044
32.242.208.160.168.102.128.121.134.044
96.197.144.109.145.092.128.121.134.044
Never.172.098.071.135.087.128.121.134.044

At short HH, recent BOLD history supports high accuracy: history-only ridge matches or exceeds full-input RR-AR, while RR-AR matches or outperforms BrainVista and outperforms MLP and GRU. At long HH, history-only accuracy declines and adding stimulus improves prediction; RR-AR remains stronger than BrainVista, MLP, and GRU. Thus, the main patterns—strong linear prediction with recent history and increasing benefit from stimulus as history becomes distant—persist on unseen stimuli in both subject cohorts. Appendix J of the revised manuscript provides the complete curves (Figure 19), numerical results (Tables 22–23), familiar-stimulus comparison, and implementation and statistical details.

The crossover is consistent with our central finding that the relative predictive value of endogenous BOLD history and exogenous stimulus information depends on measurement recency. Frequently refreshed history can be predictive enough that stimulus features add little or no gain; as history ages, stimulus information becomes more useful.


AuthorsOpenReview ↗

Response to model size and prediction terminology

General Question 3 — Parameter-matched controls (W4)

Summary. Why use parameter-matched nonlinear controls, and does the benchmark also evaluate larger nonlinear models?

Response.

Parameter matching applies only to the MLP and GRU diagnostic controls. Both receive the same stimulus features and BOLD-history window as RR-AR, with parameter counts matched within 1%. This comparison asks whether replacing the linear mapping with a nonlinear predictor improves prediction at a comparable parameter budget.

The broader benchmark includes architectures at their declared sizes, including substantially larger nonlinear predictors. The table reproduces the downstream parameter counts from the revised manuscript; frozen stimulus encoders are excluded.

ModelHCPBOLD MomentsAlgonauts
RR-AR19.00919.00919.009
MLP / GRUWithin 1% of RR-ARWithin 1% of RR-ARWithin 1% of RR-AR
TRIBE v2356.526153.675143.430
BrainVista7.6497.6497.649
NeuroWorld7.6137.6137.613
MIRAGE1,097.6791,097.6791,097.679

Counts are in millions of allocated downstream parameters or fitted coefficients, including output offsets. The matched controls compare linear and nonlinear predictors at a similar parameter budget, while the broader benchmark evaluates other architectures at their declared sizes under the shared-input protocol. Together, these comparisons address where model complexity improves prediction.

General Question 4 — Prediction-horizon terminology (W5)

Summary. How does the refresh interval relate to the established prediction horizon?

Response. We clarify the relation to the established term prediction horizon. Within each prediction block, HH is the maximum prediction horizon: the model predicts steps 1 through HH from the preceding C=8C=8 BOLD history frames, feeding each prediction into its rolling history. We use refresh interval to specify that this prediction block repeats after the next measured-history update.

For example, with H=4H=4, measured frames 1–8 initialize predictions for frames 9–12; measured frames 5–12 then initialize predictions for frames 13–16. Targets within a prediction block are never supplied as inputs. Predictions from all blocks are concatenated and scored on the same target time points for every HH; the reported score therefore summarizes the block rather than only its final HH-step-ahead prediction. The no-refresh condition uses measured BOLD only in the initial context.

The revised manuscript explains this relation in the evaluation setup and Appendix A.6 (Prediction protocol). It also distinguishes the context length CC, training rollout length KK, and test interval HH.

Reviewer_cNEfOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors analyze methods to decide naturalistic movie stimuli from BOLD measurements of human subjects. Three different established benchmark data sets are decoded with linear and nonlinear methods and different stimulus encoding approaches, pretrained neural networks are compared with raw stimuli.

The findings indicate that the more recent BOLD information is included in the feature set for decoding the next stimulus, the higher the correlation between predicted and true stimulus. This is somewhat expected, given the spatiotemporal autocorrelation structure of the hemodynamic response function (HRF).

It actually appears that there is a methodological detail that could introduce train/test leakage and thus jeopardize some aspects of some of the referenced literature and the results with autoregressive terms, meaning past BOLD activation as features. The results indicate that a refresh parameter smaller that 16 consistently outperforms other settings. The HRF would also suggest that within that time window (even with the 3 samples shift explained in the appendix) there should be mutual information between past activation and current BOLD activity.

It would be helpful if the authors could clarify the potential train/test leakage problem. Maybe a direct comparison of the information of the RR-AR control condition with pure autoregressive input would be helpful for that? There are comparisons, which is great, but quantifying the information in a way that makes it comparable across conditions would make the assessment easier, such as maybe some information theoretic metric like mutual information between BOLD and stimulus.

The exact effects here are difficult to assess and attribute to stimulus and autocorrelation structure, as there as the shift of 3 samples as described in the appendix, but then there are various other preprocessing or analysis choices that affect the cross-correlogram between stimulus and BOLD signal.

Overall the paper addresses a number of interesting research questions and with some issues regarding clarity (eg wrt the HRF correction / label leakage), the experiments are conducted well and according to scientific standards. Some details related to the neuroscientific analyses are maybe better suited for a more specialised journal, like Neuroimage, and not as interesting for a ML audience. And some details of the analysis could be controlled a bit better to enable a better disentangling of the stimulus-BOLD correlations and the BOLD autocorrelation.

Critical strengths

Interesting experimental conditions

Multiple data sets and modelling conditions

state of the art analyses in some respects (the analyses appear to match the state of the art of decoding, while some of their assumptions might be a bit over simplistic, but that's not a weakness of the paper as it's probably standard across many other decoding studies)

Critical weaknesses

Would be helpful to report dimensionality of all features and stimuli and the number of samples - the appendix is great, but maybe highlighting these simple feature/sample counts helps to get an intuition of the problem, too.

HRF is assumed constant and space-time separable in the shift that accounts for the hemodynamic lag

Potential train test leakage due to autocorrelation of HRF

Some aspects are difficult to disentangle, like for instance the amount of information in the current BOLD signal and past BOLD signal and the stimulus. There are some interesting analyses also in the simulation experiments, but the refresh rate parameter as well as the constant lag and assumption of a constant HRF renders it difficult to relate the effects to hemodynamic or preprocessing.

Questions for authors

It would be great if there were a quantification of the entropy in the BOLD related to the stimulus and prior BOLD signal across all conditions

Can the authors exclude leakage of train data in the test window due to the HRF dynamics?

4 comments
AuthorsOpenReview ↗

Response to benchmark scope, data separation, and ML relevance

We thank the reviewer for these comments and would first like to clarify that the prediction target in our study is BOLD activity. History-conditioned predictors use stimulus features and BOLD history as model inputs to predict subsequent BOLD responses.

General Question 1 — Feature and sample counts (W1)

Summary. What are the feature dimensions and sample counts?

Response.

We summarize the feature dimensions and sample counts below to make the scale of the shared-input benchmark explicit.

Input or targetDimensions per time point
Video: V-JEPA2 ViT-g2×1,408=2,816
Audio: Wav2Vec-BERT 2.02×1,024=2,048
Text: Llama 3.2-3B2×3,072=6,144
Concatenated video/audio/text11,008
BOLD target (Schaefer parcels)1,000
Raw RGB pixels (visual-input control)16×16×3=768

With C=8C=8 BOLD history frames (8,0008{,}000 values), the full RR-AR model maps 19,00819{,}008 input coordinates (11,00811{,}008 stimulus + 8,0008{,}000 history) to 1,0001{,}000 BOLD targets.

Each entry below gives train / validation / test counts per fold.

DatasetSubjectsRecordsTarget time points
HCP107 / 15 / 321,926 / 270 / 576343,898 / 48,210 / 102,848
BOLD Moments8 / 8 / 288 / 8 / 2422,616 / 2,056 / 6,168
Algonauts3 / 3 / 1927 / 117 / 348417,369 / 52,323 / 156,564

HCP uses one fixed subject split; BOLD Moments uses five held-out subject-pair folds; Algonauts uses four leave-one-subject-out folds. Training and validation in the latter two datasets use the same subjects but separate scans or stimulus groups. Records are movie-clip presentations, continuous scans, and movie segments, respectively.

General Question 2 — Temporal dependence and leakage (W3 / Q2 / S1 / S2)

Summary. Does using past BOLD introduce train–test leakage?

Response. Training and test recordings come from disjoint subjects, rather than adjacent portions of the same BOLD time series. HRF-induced temporal dependence within a held-out subject’s recording therefore does not transfer training observations into the test window. Past BOLD from that subject is an explicitly permitted input in our history-conditioned prediction task.

At each step, a history-conditioned model predicts yt\mathbf y_t from the preceding C=8C=8 history frames available under the evaluation protocol, together with stimulus inputs when included. The current target yt\mathbf y_t and future measured BOLD frames are not supplied as history inputs. Test subjects are excluded from model fitting and validation-based selection, and model parameters remain fixed throughout testing.

General Question 5 — Relevance to machine learning (S4)

Summary. What is the ML contribution and its relevance to ICLR?

Response. Thank you for pointing this out. ICLR’s reviewer guidelines explicitly recognize new, relevant, and impactful knowledge, including empirical findings, as a contribution. Methodological novelty is therefore not the only form of contribution recognized by ICLR.

Our study contributes new knowledge through three empirical insights:

  1. Representation quality and predictor complexity provide different benefits. Pretrained stimulus representations generally improve prediction, while increasing downstream complexity does not consistently improve performance under matched inputs.
  2. The value of model complexity depends on available history. Recent BOLD supports strong linear prediction, while stimulus input becomes more valuable as that history grows distant.
  3. The observation process shapes predictability. Paired simulations show that slower hemodynamic responses can prolong history-based predictability, including under nonlinear neural dynamics.

These findings inform where to allocate model complexity and support comparing predictors under matched access to measured BOLD history. This contribution is consistent with empirical representation studies published at ICLR.

Our study complements these representation studies by examining when additional predictor complexity improves BOLD prediction under matched stimulus representations and history availability.

AuthorsOpenReview ↗

Response to HRF approximation and interpretation

General Question 4 — HRF approximation and interpretation (W2 / W4 / S3 / S4)

Summary. Why use a fixed-delay HRF approximation, and how do the controls support a hemodynamic interpretation?

Response.

(1) Why approximate the HRF with a fixed delay? (W2) A fixed shift captures temporal alignment but simplifies the HRF's shape and variability. Estimating neural responses jointly with region- and subject-varying HRFs is difficult. We therefore follow the fixed-delay convention used by TRIBE, MIRAGE, and NeuroWorld, providing a common alignment for model comparison (Appendix A.4). We stress-tested this alignment choice using four delays. Across the tested delays, RR-AR retains higher prediction accuracy than MLP and RR-stim, and the pattern of strong prediction with recent BOLD history and increasing benefit from stimulus as history grows distant persists (Appendix D.7, Tables 9–10).

(2) How do we identify the hemodynamic contribution? (W4) Recent BOLD supports strong short-interval linear prediction. Because BOLD reflects a hemodynamic transformation of neural activity, its predictability also depends on this observation process. A slower HRF may prolong temporal dependence, keeping history informative for longer. We therefore test whether changing the HRF changes predictability while holding neural activity fixed, using four controls:

  • Same neural trajectories: slower HRFs sustain high accuracy for longer; the no-HRF reference lacks the short-interval plateau (Section 4.4; Appendix F).
  • Same history availability: the pattern persists when HRFs are compared at identical refresh intervals, with matched stimuli, splits, and targets (Appendix G).
  • No stimulus input: history-only predictors retain the pattern, separating it from stimulus-feature alignment (Appendix G).
  • Nonlinear neural dynamics: slower HRFs retain greater RR-AR predictability at H=32H=32 across the tested schedules and nonlinear strengths (Appendix F.2).

In these simulations, the matched comparisons identify a contribution of the hemodynamic time scale to history-based predictability.

In the empirical benchmark, we also varied the history context (1, 4, or 8 BOLD frames) to assess sensitivity to the amount of available BOLD history. RR-AR remains competitive, while stimulus-only models can overtake it at long refresh intervals (Appendices D.1–D.2).

AuthorsOpenReview ↗

Response to information from history and stimulus

General Question 3 — Information supplied by history and stimulus (W4 / Q1 / S2 (information))

Summary. How much information about current BOLD is associated with stimulus and past BOLD across conditions?

Response.

(1) Prediction controls: what do history and stimulus add? At short refresh intervals (H≤4H\leq4), RR-AR, BrainVista, and MLP outperform all three stimulus-only references across all three datasets. History-only RR-AR nearly matches the full model at these intervals. As refresh intervals lengthen, history-only accuracy declines while stimulus-only predictions remain unchanged. Adding stimulus yields larger gains over history-only RR-AR on HCP and Algonauts than on BOLD Moments at long intervals. Figure 12 reports the complete comparison of full-input, history-only, and stimulus-only predictors.

(2) Quantifying information from each source. For each held-out subject and target parcel, YY is current BOLD, BB is that parcel's eight-frame earlier BOLD history, and SS comprises the first 16 principal components of aligned stimulus features, with PCA fitted on training data. Using uncorrected Gaussian-copula estimation (Appendix H.2), we estimate I(Y;S)I(Y;S), I(Y;B)I(Y;B), I(Y;S∣B)I(Y;S\mid B), and I(Y;B∣S)I(Y;B\mid S) in bits. The conditional quantities measure the reduction in uncertainty about YY associated with one source given the other; marginal information values can overlap.

The analysis includes all 1,000 Schaefer parcels in every held-out subject: 32 HCP subjects, 10 BOLD Moments subjects across five held-out-pair folds, and four Algonauts subjects across four leave-one-subject-out folds. Table values average parcel estimates within each subject, then equally across subjects (Appendix H.3).

(3) Information changes with history distance. The tables cover all nine tested distances. Here, kk is the separation in TRs between the target and the most recent history frame; HH is the refresh interval used during prediction. Target time points and stimulus descriptors are fixed across kk, so I(Y;S)I(Y;S) is unchanged.

HCP

Information / k124816326496128
I(Y;S)0.047790.047790.047790.047790.047790.047790.047790.047790.04779
I(Y;B)0.534560.291290.076490.016420.009460.009290.007330.007380.00702
I(Y;S∣ B)0.014600.023390.036370.046520.049470.048490.047870.047980.04794
I(Y;B∣ S)0.501360.266890.065060.015150.011130.009990.007400.007560.00716

BOLD Moments

Information / k124816326496128
I(Y;S)0.011830.011830.011830.011830.011830.011830.011830.011830.01183
I(Y;B)0.151000.039160.020350.022140.011020.004620.004460.004450.00461
I(Y;S∣ B)0.009630.011450.011780.011760.011830.011830.011900.011840.01189
I(Y;B∣ S)0.148800.038780.020300.022080.011020.004620.004530.004470.00468

Algonauts

Information / k124816326496128
I(Y;S)0.028820.028820.028820.028820.028820.028820.028820.028820.02882
I(Y;B)0.229370.084690.016200.017240.012070.001430.000800.000120.00012
I(Y;S∣ B)0.011700.021590.028290.028690.028040.028630.028750.028830.02881
I(Y;B∣ S)0.212240.077450.015660.017100.011280.001230.000730.000130.00011

Recent local BOLD history remains informative after conditioning on stimulus. At distant history lags, history information is lower and stimulus provides more additional information, consistent with the prediction controls.

Thank you again for your thoughtful review. As the discussion period draws to a close, we would appreciate it if you could take our additional results and analyses into account in your final assessment. If any questions remain, we would be happy to clarify them.

Reviewer_6GytOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies when model complexity improves cross-subject prediction of BOLD signals during naturalistic stimulation. It separates the complexity of stimulus representations from that of downstream predictors and compares raw pixels against pretrained stimulus features, and linear ridge models against nonlinear stimulus-only and history-conditioned predictors. Eight benchmark predictors are assessed on HCP movie fMRI (154 subjects), BOLD Moments (10 subjects), and Algonauts 2025 (4 subjects), with held-out subjects and no subject-specific refitting. The results show that pretrained visual features generally improve prediction, recent measured BOLD history yields substantial gains that diminish with longer refresh intervals, and ridge autoregression is a strong reference for history-conditioned prediction. Regional consistency analyses, functional-connectivity comparisons, and controlled hemodynamic simulations further characterize where and why history helps.

Critical strengths

Strength 1: The benchmark separates representation quality from downstream complexity and tests eight predictors across three naturalistic fMRI datasets. The controlled refresh-interval protocol, parameter-matched MLP/GRU references, ridge-fitting controls, and reduced-rank analyses make the comparison informative and practically useful.

Strength 2: The paper goes beyond a model leaderboard: the regional response-consistency and functional-connectivity analyses identify where measured history adds value, while fixed-neural-trajectory hemodynamic simulations provide a concrete explanation for longer-lasting predictability under slower simulated responses.

Critical weaknesses

Weakness 1: The main architectural conclusion is sensitive to benchmark adaptation and training. MIRAGE replaces its native multimodal representations and subject-specific heads, while BrainVista and NeuroWorld are benchmark reimplementations; downstream neural models are trained from scratch on common cached inputs. Consequently, RR-AR outperforming these implementations does not establish that linear autoregression would outperform the original complete nonlinear systems. The paper acknowledges this scope, but the central framing sometimes invites a broader interpretation.

Weakness 2: The evaluation primarily tests unseen subjects viewing familiar stimuli: Table 1 reports 100% train–test stimulus overlap for HCP and BOLD Moments and 89.7% for Algonauts. Moreover, history-conditioned methods receive periodic measured target-subject BOLD updates; they therefore solve a materially different information-availability problem from stimulus-only prediction. These design choices are legitimate and described, but they restrict conclusions about unseen-stimulus transfer and deployment without frequent measurements.

Questions for authors

Clarifications

  1. Can the authors quantify the extent to which the key RR-AR versus neural-model rankings depend on benchmark-specific changes? For example, an ablation that preserves more of MIRAGE’s native representations/readout, or a focused analysis of training objectives and optimization differences for the BrainVista/NeuroWorld adaptations, would help distinguish architectural effects from implementation effects.

  2. Do the main patterns persist under a stimulus-disjoint cross-subject evaluation on at least one dataset? If such a split is infeasible, please quantify its constraints and clarify which conclusions are specific to held-out participants viewing familiar content. Please also explain how complete-recording offline preprocessing affects the interpretation of future online or strictly causal prediction.

Suggestions for future work Evaluate joint subject-and-stimulus generalization and strictly causal online pipelines, including the cost and feasibility of obtaining periodic measured BOLD refreshes.

4 comments
AuthorsOpenReview ↗

Response to official implementations

General Question 1 — Official model implementations (W1 / Q1)

Summary. How do benchmark adaptations affect the comparison between RR-AR and neural models, and do the findings persist when official model components are restored?

Response.

(1) Why adapt the models for this benchmark? We appreciate the reviewer's concern about the influence of implementation choices. Our original benchmark was designed to compare downstream predictors under common stimulus inputs and matched BOLD-history availability. For MIRAGE, we therefore retained the released downstream network while replacing its Qwen features with the common features used by the other predictors. For BrainVista, whose official code was unavailable when we implemented the benchmark, we implemented the model from the published description and integrated it into the same evaluation framework.

(2) Additional runs using the authors' released implementations. For MIRAGE, we restore native Qwen3-Omni representations and released model components, train the downstream predictor from scratch on training subjects, and average their readout heads for unseen subjects. For BrainVista, we replace our paper-based implementation with the released model components and official training procedure. Both evaluations retain our cross-subject protocol.

The additional BrainVista results cover all three datasets, and MIRAGE covers HCP and BOLD Moments. MIRAGE on Algonauts would require more than two weeks of feature extraction on our available hardware, exceeding the revision window.

(3) How do the official-code models compare with RR-AR? The tables compare manuscript (M) and official-code (O) results against RR-AR. Rows show H (TRs); entries are mean (subject SD). MIRAGE is stimulus-only, so its score repeats across H; a dash indicates an unavailable result.

HCP

HRR-ARBrainVista MBrainVista OMIRAGE MMIRAGE O
1.685(.032).618(.035).551(.028).267(.049).047(.013)
2.632(.031).563(.034).506(.029).267(.049).047(.013)
4.541(.033).472(.034).424(.031).267(.049).047(.013)
8.444(.033).377(.030).332(.032).267(.049).047(.013)
16.377(.034).310(.025).264(.030).267(.049).047(.013)
32.336(.037).270(.024).224(.031).267(.049).047(.013)
96.304(.045).241(.031).188(.032).267(.049).047(.013)
Never.294(.049).217(.031).178(.032).267(.049).047(.013)

BOLD Moments

HRR-ARBrainVista MBrainVista OMIRAGE MMIRAGE O
1.408(.039).305(.043).320(.028).090(.008).104(.012)
2.329(.032).227(.036).244(.024).090(.008).104(.012)
4.254(.021).156(.025).173(.016).090(.008).104(.012)
8.193(.015).109(.020).127(.011).090(.008).104(.012)
16.147(.010).082(.014).097(.007).090(.008).104(.012)
32.110(.008).058(.010).079(.008).090(.008).104(.012)
96.073(.008).029(.008).064(.008).090(.008).104(.012)
Never.054(.009).002(.001).058(.008).090(.008).104(.012)

Algonauts

HRR-ARBrainVista MBrainVista OMIRAGE MMIRAGE O
1.514(.033).438(.028).424(.030).240(.011)—
2.449(.024).356(.015).353(.021).240(.011)—
4.368(.016).259(.008).265(.012).240(.011)—
8.308(.013).193(.007).201(.010).240(.011)—
16.268(.011).152(.008).161(.011).240(.011)—
32.241(.011).125(.011).140(.014).240(.011)—
96.222(.011).096(.016).127(.018).240(.011)—
Never.212(.012).022(.008).136(.020).240(.011)—

Comparison with RR-AR. The official-code reruns retain RR-AR's advantage over BrainVista on HCP and Algonauts. On BOLD Moments, RR-AR performs better at short intervals, while the two models become close at long intervals and nearly match without refresh (0.054 versus 0.058). Official-code MIRAGE remains well below RR-AR on HCP; on BOLD Moments, it is close to RR-AR at H=32H=32 (0.104 versus 0.110) and performs better at H=96H=96 and without refresh. Overall, RR-AR remains a strong reference when recent measured history is available, while the benefit of nonlinear stimulus-based prediction becomes more apparent as that history grows distant.

An HCP audit found no scoring or duplicated-offset error. Head averaging reduced accuracy in a training diagnostic, but scores were low with both readouts; it is therefore only a possible contributor to the low score (Appendix K.4).

Implementation and evaluation details, aggregation procedures, and complete results are provided in Appendices B and K.4.


AuthorsOpenReview ↗

Response to generalization, online prediction, and future work

General Question 2 — Unseen-stimulus generalization and online prediction (W2 / Q2)

Summary. Do the main findings persist for unseen subjects viewing unseen stimuli, and how do measured-history availability and offline fMRI preprocessing affect their interpretation for online prediction?

Response.

(1) Generalization to unseen stimuli. We added a stimulus-disjoint HCP evaluation: models are trained on Movie 1/2, selected on Movie 3, and tested on Movie 4, with repeated clips excluded across partitions. We test both previously seen subjects and entirely held-out subjects. The comparisons include RR-AR, history-only ridge, RR-stim, MLP, GRU, BrainVista, TRIBE v2, and both shared-input and official-code MIRAGE configurations.

History-conditioned models use preceding available BOLD observations, whereas stimulus-only models receive no measured BOLD. Representative mean correlations for held-out subjects viewing unseen stimuli are shown below. Stimulus-only scores are independent of HH.

HCP: unseen subjects and unseen stimuli

Hist. = history-only ridge; BV = BrainVista; S/O = shared-input/official-code. BV and TRIBE v2 use shared inputs.

HRR-ARHist.BVMLPGRURR-stimTRIBE v2MIRAGE SMIRAGE O
1.621.680.621.386.210.128.121.134.044
2.569.622.555.378.205.128.121.134.044
4.479.515.440.316.176.128.121.134.044
8.378.387.318.249.141.128.121.134.044
16.296.283.227.201.115.128.121.134.044
32.242.208.160.168.102.128.121.134.044
96.197.144.109.145.092.128.121.134.044
Never.172.098.071.135.087.128.121.134.044

At short HH, recent BOLD history supports high accuracy: history-only ridge matches or exceeds full-input RR-AR, while RR-AR matches or outperforms BrainVista and outperforms MLP and GRU. At long HH, history-only accuracy declines and adding stimulus improves prediction; RR-AR remains stronger than BrainVista, MLP, and GRU. Thus, the main patterns—strong linear prediction with recent history and increasing benefit from stimulus as history becomes distant—persist on unseen stimuli in both subject cohorts. Appendix J of the revised manuscript provides the complete curves (Figure 19), numerical results (Tables 22–23), familiar-stimulus comparison, and implementation and statistical details.

(2) Offline fMRI preprocessing, stimulus feature extraction, and online prediction. We distinguish fMRI preprocessing from stimulus feature extraction:

  • fMRI preprocessing—BOLD Moments: nuisance regression and detrending are fitted to each complete fMRI scan, followed by within-scan standardization of the BOLD time series. These coefficients and statistics use later BOLD samples from the same scan, including held-out test scans.
  • fMRI preprocessing—Algonauts and HCP: Algonauts retains the source standardization of each complete fMRI time series, using its full-duration mean and standard deviation. HCP uses supplied hp2000-clean fMRI data; our parcel extraction adds no temporal filtering or per-scan scaling.
  • Stimulus feature extraction: video, audio, and text features are extracted offline before predictor fitting.

There are direct precedents for these choices: BrainVista (Section 3.3) uses run-wise temporal z-scoring of fMRI alongside causal autoregressive prediction; TRIBE (Sections 2.2–2.3) uses session-wise fMRI z-scoring. Our results concern prediction of offline-preprocessed BOLD using offline-extracted stimulus features. For strictly online use, fMRI nuisance/scaling estimates would need to use calibration or past fMRI observations.

General Question 3 — Suggestions for future work (F1)

Summary. Extend the evaluation to joint subject-and-stimulus generalization and strictly causal online prediction, including the practical availability of measured BOLD history.

Response. We have now added a joint held-out-subject and unseen-stimulus HCP evaluation (Appendix J). Future work will address strictly causal online pipelines and the practical cost and feasibility of obtaining periodic measured BOLD updates.

Thank you again for your thoughtful review. As the discussion period draws to a close, we would appreciate it if you could take our additional results and analyses into account in your final assessment. If any questions remain, we would be happy to clarify them.

Thank you for your deeper and complementary analysis of generalizations, notably regarding the benchmark training. It's more compelling than it was initially.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to AC

Dear Area Chair,

Thank you for overseeing the review of our submission. We have revised the manuscript and provided detailed point-by-point responses to all four reviewers, including substantial additional experiments, controls, and statistical analyses.

We would like to briefly highlight the key additions:

  1. Generalization to unseen stimuli and additional simple baselines (Reviewers 6Gyt, VAXd, pp2X). We added a stimulus-disjoint HCP evaluation, training on Movies 1–2, selecting models on Movie 3, and testing on Movie 4, with repeated clips excluded across partitions. The evaluation covers both previously seen and entirely held-out subjects. Strong linear prediction with recent BOLD history and increasing benefit from stimulus information as history becomes distant persist in both cohorts. We also evaluated the requested training-subject mean and exact history k-NN baselines on familiar stimuli; neither reproduces RR-AR’s accuracy on the same held-out subjects and prediction targets.
  2. Additional evaluations using official model implementations (Reviewers 6Gyt, pp2X). We added BrainVista evaluations using released model components and the official training procedure across all three datasets, and MIRAGE evaluations restoring native Qwen3-Omni representations and released components on HCP and BOLD Moments. These evaluations retain our cross-subject protocol. RR-AR remains a strong reference when recent measured history is available. We also explicitly report exceptions: on BOLD Moments, official-code MIRAGE outperforms RR-AR at long refresh intervals, while official-code BrainVista nearly matches it without refresh. These results refine the conditions under which additional model complexity is beneficial.
  3. Disentangling history and stimulus contributions (Reviewers cNEf, pp2X). We added Gaussian-copula estimates of marginal and conditional information across all three datasets, complementing the history-only prediction controls. Recent BOLD history remains informative after conditioning on stimulus; as history becomes more distant, its information decreases and stimulus provides more additional information. We also tested four stimulus–BOLD alignment delays, retaining the main predictive patterns. These analyses clarify how measurement recency affects the relative contributions of history and stimulus.
  4. Extension to nonlinear neural dynamics (Reviewers cNEf, pp2X). We extended the paired simulations to two strengths of nonlinear neural coupling, comparing different hemodynamic responses applied to identical neural trajectories. Slower responses continue to sustain strong linear predictability over longer intervals, including in history-only controls. RR-AR performs strongly at short and intermediate intervals, while nonlinear predictors can match or outperform it in some conditions without further measured-history updates. This supports a role for the observation time scale without implying universal superiority of linear prediction.
  5. Stronger statistical evidence and numerical reporting (Reviewer pp2X). We added complete numerical results underlying the figures and subject-level paired comparisons with pointwise 95% confidence intervals and Holm-adjusted p-values. We explicitly distinguish independent subject counts from prediction time points and acknowledge the limitations of the smaller cohorts. In the shared-input benchmark, RR-AR outperforms parameter-matched MLP and GRU at refresh intervals of 1, 2, 4, and 8 TRs across all three datasets after correction. We have also clarified the subject-disjoint data separation, permitted use of preceding BOLD observations, offline preprocessing assumptions, and relationship between refresh interval and prediction horizon. Parameter matching applies only to the MLP/GRU controls; the broader benchmark includes substantially larger nonlinear models.

Together, these additions strengthen the paper’s empirical contribution: separating the benefits of stimulus representation from downstream predictor complexity, and identifying how recent measurement availability and the hemodynamic observation process shape their predictive value.

We hope these revisions address the main concerns raised during review. We would greatly appreciate any further discussion while there is still time to address remaining questions before November 18.

Thank you again for your time and consideration.

Best regards, Authors

Dear reviewers,

The rebuttal period is drawing to an end. Please confirm that you have read authors rebuttals and either acknowledge completeness or ask your remaining questions.

your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.