The work
behind the claim.
BrainVI's research grounds everything we build. These are early-stage preprints and whitepapers — single-authored, openly available, and written to report limits as plainly as results. We report the noise ceiling next to every number.
- ↘ fsaverage6 · 81,924 vertices
- ↘ public CC0 / CC-BY datasets
- ↘ reproductions on vetted request, post model release
2026
The Average Brain Is No Brain At All: A Zero-Shot Evaluation of TRIBE v2 on Out-of-Distribution Naturalistic Video
A clinical evaluation of Meta's publicly released TRIBE v2 brain-encoding checkpoint. We show that its average-subject embedding, applied without per-subject adaptation, captures only ~4.3% of the measured inter-subject noise ceiling — and is actively anti-correlated in early visual cortex, the region with the highest reliability. A Yeo-7 network decomposition reveals a clean hierarchy: higher-order association cortex transfers across subjects far better than sensory cortex. The result establishes why subject-specific adaptation, not raw backbone scale, is the dominant factor in brain prediction.
Read PDF→2026
MARY-Nano: A Six-Stream Multimodal Brain Encoder for In-Silico Neural Prediction
MARY-Nano is a ~35M-trainable-parameter adapter over six frozen foundation-model backbones — one per sensory and semantic pathway (motion, vision-language, audio, speech, long-context language, and on-screen text) — that predicts cortical activity natively across the 81,924-vertex fsaverage6 surface, the highest-resolution brain-encoding output to date (4× the field's fsaverage5). Trained on just ~23 hours of CNeuroMod Friends fMRI — roughly a third of TRIBE v2's data — MARY-Nano 1.0 matches the Algonauts 2025 state of the art on the field's Schaefer-1000 benchmark: out-of-distribution Pearson r = 0.216 vs TRIBE's 0.2146. On held-out films it generalizes at r = 0.170, and its SRM head recovers novel-subject prediction 4–5× in minutes with no gradient training. Scaling the training set by two films (MARY-Nano 1.1) lifts held-out generalization a further +9% to 0.185 — confirming film diversity, not model size, as the dominant lever.
Read PDF→2026
Cross-Modal Neural Translation via Synthetic fMRI
Can computationally predicted fMRI drive the same decoders trained on real fMRI? We encode a single modality into synthetic cortical activity, then decode it back into images and text via MindEye2 and into audio via AudioLDM2 — six directional pipelines through one neural bottleneck. In a small proof-of-concept (n = 20 test images), synthetic fMRI recovered CLIP similarity of 0.737 versus 0.749 for real fMRI — within noise at this sample size, and a directional signal rather than a validated result. A 2×2 ablation finds that training-data scale, not adapter capacity, is the dominant lever.
Read PDF→2026
Retinotopic Decoding from Synthetic fMRI: Mapping Predicted Cortical Activity to Visual Field Images
Using the Benson 2014 retinotopic atlas, we project predicted cortical activity in V1–V3 back into visual-field coordinates and render it via Gaussian splatting — no generative model, only atlas geometry. We first validate the pipeline on ground-truth 7T fMRI from the Natural Scenes Dataset, where decoded visual fields capture coarse spatial structure (object boundaries, figure-ground contrast) directly from measured brain responses, then apply the same readout to model predictions. A transparent diagnostic for whether an encoder has learned physiologically plausible spatial structure.
Read PDF→Open,
verifiable.
Our papers are single-authored preprints and whitepapers — not peer-reviewed venue publications. We will provide traces and reproductions for verification by the research community on vetted request, following our model release. We believe open verification is how trust is earned and how the field moves forward.