Generalization of brain–model alignment metrics to other architectures and objectives

Establish whether comparable spatial, temporal, and overall encoding alignment metrics between model representations and human brain activity would emerge when using architectures and training objectives other than the self-supervised, hierarchical DINOv3 vision transformer family evaluated in this study.

Background

The paper measures representational alignment between DINOv3 vision transformers and human brain activity using three metrics: an overall encoding score (linear predictability), a spatial score (hierarchical layer-to-cortical distance correspondence in fMRI), and a temporal score (layer-to-MEG response latency correspondence).

All results are based on a single family of self-supervised, hierarchical computer vision models (DINOv3). The authors therefore identify uncertainty about whether the observed alignment patterns would hold across different neural network architectures and training paradigms.

References

It thus remains an open question whether similar spatial, temporal and encoding scores would emerge with other architectures and training objectives.

— Disentangling the Factors of Convergence between Brains and Computer Vision Models  (2508.18226 - Raugel et al., 25 Aug 2025) in Discussion — Limitations

As a proof-of-concept, we demonstrated our framework using a self-supervised contrastive ResNet-50 model and evaluated alignment at one model layer using a single neural benchmark. Whether the observed trends generalize across other architectures, learning objectives, model layers, and datasets remains to be determined.

— Bidirectional representational alignment between biological and artificial neural networks  (2608.18244 - Kostousov et al., 18 Aug 2026) in Discussion, paragraph beginning “One limitation of our work”

We restrict our evaluation to these three models, spanning a coarse robustness spectrum; we have not verified whether the trends reported here hold for other architectures (e.g., other ViT scales, other self-supervised objectives, or other vision-LLMs), and we do not claim this small set is representative of the full robustness spectrum.

— Technical note on: Zero-Training Feature-Space Alignment via Information Geometry  (2609.37302 - Khan et al., 29 Sep 2026) in Section 4.1, Experimental Setup, Models

Yet representational similarity alone does not establish functional redundancy, and whether these patterns extend to 3DVGTs remains unexplored.

— Does the VGGT Family Need All Its Layers?  (2609.36842 - Zhang et al., 29 Sep 2026) in Appendix A.2, Layer Redundancy in Transformers

We study one LLM (LLaMA2-7B, one intermediate layer), one fMRI dataset, and seven participants; our conclusions are specific to this setting, and sensitivity to the choice of layer and generalization to newer or larger LLMs, whose brain alignment varies with scale, remain untested.

— Decomposing Brain-LLM Alignment with Interpretable Semantic Representations  (2610.21508 - Kolling et al., 15 Oct 2026) in Limitations, Section 5 (Conclusion)