- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes CORTIVA, a candidate-score fusion framework for EEG- and MEG-to-image retrieval. It uses three neural decoding routes (SAR, CVR, and EPR) aligned with different visual targets. Instead of combining their embeddings before retrieval, the method computes candidate scores separately and then combines them through weighted summation. The authors report 73.5% Top-1 accuracy on THINGS-EEG2 and 42.4% on THINGS-MEG under within-participant evaluation. The paper also includes cross-participant adaptation, route-removal ablations, weight controls, and several neural and visual representation analyses. Although the multi-route design has some practical value and the authors report substantial experiments with positive results, the paper does not sufficiently establish the motivation for its research question, its distinct methodological contribution, or how directly the experiments support the claimed benefit of score-level fusion. Together with incomplete comparisons and reproducibility concerns, these issues weaken the overall scientific contribution of this work.
Critical strengths
The main strength of this paper is its relatively careful empirical evaluation, particularly the effort to examine the contribution of individual routes rather than relying only on final retrieval accuracy. The authors retrain route-removal variants under matched settings across ten EEG participants and five seeds, showing that removing CVR and EPR reduces Top-1 accuracy by 5.47 and 5.40 percentage points, respectively, with positive simultaneous 95% confidence intervals (Figure 3; Table S15). The four weighting controls are also conducted on the same 50 participant-by-seed score matrices, providing useful evidence that the benefit of combining route scores remains under uniform weighting (Figure 7; Table S16). I also appreciate the participant-level multi-seed evaluation, negative controls, and independent DINOv2 analyses, which provide additional evidence for the reported retrieval behavior (Tables S6, S13, S20–S21).
Critical weaknesses
-
The research question is not sufficiently clear or well motivated. This is the main concern. The Abstract and Introduction argue that early embedding consolidation loses disagreement among visual targets, but they do not convincingly explain why this is an important problem for EEG/MEG retrieval. What kind of disagreement is actually useful, and under what conditions does early consolidation hurt retrieval? The reader is largely left to infer the specific research question from the proposed architecture. This goes beyond a writing issue, as the motivation for the central contribution is not sufficiently established. (Abstract and Section 1, pp. 1–2)
-
The methodological novelty remains unclear. Combining scores from multiple retrieval routes is closely related to established late-fusion and ensemble methods. The visual backbones and many training components are also existing techniques. Using existing components is not a weakness by itself. Their selection and combination should, however, be clearly motivated by the research problem and supported by evidence showing why these design choices are appropriate. This connection is not sufficiently established in the paper. Moreover, CVR and EPR already combine visual features or projections internally, while the paper argues for preserving disagreement until score fusion. The rationale for allowing early consolidation within these routes but not across them remains unclear. (Section 2, p. 2; Figure 1, p. 4; Appendix B.4, p. 16; Table S7, p. 32)
-
The experiments do not sufficiently isolate the claimed advantage of score-level fusion. The matched route-removal experiments provide useful evidence for the contribution of individual routes, but removing a route also changes visual supervision, model components, and training objectives. These results do not directly establish that score-level fusion is better than embedding-level fusion under comparable conditions. A matched early-fusion baseline and a homogeneous multi-model ensemble are important missing controls. In addition, the query-dependent weighting rule shows no clear advantage over uniform weighting on the same trained route scores. The current evidence therefore does not sufficiently identify the source of the claimed fusion gains. (Figure 3, p. 6; Tables S14–S16, pp. 37–39)
-
Incomplete prior-art positioning and unclear protocol comparability. The closely related ICLR 2026 work Learning Brain Representation with Hierarchical Visual Embeddings is not discussed. The paper also cites Jiang et al. (2026), namely SAMGA, but does not clearly explain the methodological differences and reported results. CORTIVA does not necessarily need to outperform every recent method, particularly under different protocols, but its contribution should be more clearly positioned against these works. Furthermore, the reported cross-participant results use labeled target-participant adaptation, while the MEG evaluation uses within-participant training, 12 averaged test repetitions, and a different neural encoder. The comparability with existing baselines and the specific contribution of fusion under these settings need clearer evidence. (Section 2, p. 2; Tables 1–2, pp. 5–6; Appendix B.8–B.9, p. 19; Table S1, pp. 26–27)
-
Limited reproducibility of key implementation details. The supplementary package provides training and fusion components, configurations, and result tables, but the actual SAR/EPR backbone implementations and the scripts for producing and selecting the Stage 1 reference models are not included. The released scripts also do not provide a complete execution path from these components to the reported retrieval results. Although checkpoint manifests and validation summaries are provided, the available materials do not fully support independent verification of the model initialization and selection history underlying the reported experiments. This limits the reproducibility and traceability of the main results. (Appendix B.2–B.5, pp. 15–18; supplementary code)
Questions for authors
-
Could the authors clarify the precise scientific question behind CORTIVA? In particular, why is disagreement among visual targets important for EEG/MEG retrieval, and what evidence shows that early embedding consolidation loses useful information?
-
What is the main methodological insight beyond combining several pretrained retrieval routes? Also, why are visual representations consolidated within CVR and EPR, but preserved separately across the three final routes?
-
Could the authors further explain the advantage of score-level fusion over embedding-level fusion under comparable conditions? A matched early-fusion baseline or a homogeneous multi-model ensemble would help clarify this point. Also, what additional benefit does the proposed router provide over uniform weighting?
-
Could the authors clarify the differences between CORTIVA and the ICLR 2026 work Learning Brain Representation with Hierarchical Visual Embeddings, as well as SAMGA (Jiang et al., 2026)? In particular, how do their methodological contributions and evaluation protocols differ?
-
Are the cross-participant baselines evaluated with comparable access to labeled target-participant data? For THINGS-MEG, could the authors further explain how the contribution of score fusion is distinguished from the changes in neural encoder and evaluation settings?
-
Could the authors clarify the specific SAR/EPR backbone implementations, the provenance and selection records of the Stage 1 checkpoints, and how these checkpoints correspond to the final reported experiments? Also, the paper specifies mid-ranks for ties, while the released ranking implementation appears to assign integer ranks. Could the authors clarify which rule was used for the reported metrics and whether this difference has any practical effect? These details would help improve the reproducibility of CORTIVA.