Only Brains Align with Brains: Cross-Region Alignment Patterns Expose Limits of Normative Models
Even top-ranked models miss stable cross-region brain alignment patterns.
A new study from the University of Tübingen, presented at ICLR 2026, challenges how we evaluate AI vision models against human brains. Led by Larissa Höfling and Matthias Bethge, the team introduced a method called alignment pattern analysis (APA). Instead of just measuring how well a model predicts neural activity in one brain region, APA checks whether the model reproduces the characteristic alignment profile of each region relative to all others. Applying APA to the BOLD Moments video fMRI dataset, the researchers found that while these cross-region patterns are highly stable across different human subjects, even top-ranked models often fail to capture them.
This work exposes a critical flaw in conventional model-brain alignment benchmarks: they lack discriminative power, making many models appear equally brain-aligned when they are not. The authors argue for a clearer distinction between models that serve as predictive tools and those that claim computational or algorithmic similarity to the brain. For the latter, reproducing relational alignment patterns is a stronger and necessary test. The findings have direct implications for neuroscience and AI, suggesting that current rankings may overstate how closely AI vision systems mimic biological vision.
- APA tests whether models reproduce stable cross-region alignment profiles across human subjects.
- Conventional benchmarks lack discriminative power, making diverse models appear equivalent in brain alignment.
- Top-ranked models often fail APA, undermining claims of computational similarity to the brain.
Why It Matters
Rethinks how we validate AI vision models as brain-like, demanding stricter evidence for computational similarity.