Brain & AI · April 2026 · Meta FAIR — Brain & AI team
TRIBE: predicting whole-brain fMRI from video, audio, and text
Meta FAIR's tri-modal brain encoder won Algonauts 2025 by aligning pretrained model features to neural activity. Notes on why a predictive model of the brain matters for agent design.
Most of our work assumes the human stays in the loop — supervising, correcting, approving. So we pay close attention to research that makes the human side of that loop legible. TRIBE is one of those: a model that predicts how a brain responds to what it sees and hears, without an actual scan.
What the paper does
TRIBE (TRImodal Brain Encoder) is a ~1B-parameter model from Meta FAIR's Brain & AI team that predicts whole-brain fMRI responses to naturalistic stimuli — video, audio, and text together. It works by aligning the latent representations of pretrained foundation models (one per modality) and using a transformer to handle how those representations evolve over time, then mapping them onto neural activity.
It won the Algonauts 2025 brain-encoding competition by a wide margin, and the v2 release pushes spatial resolution roughly 70× beyond the prior state of the art, trained on hundreds of hours of fMRI and evaluated across a much larger participant set.
Why it caught my attention
Two things.
It's a predictive model, not a decoder. TRIBE simulates a brain's response to a stimulus — it does not read thoughts out of a brain. That distinction matters for how this kind of work should be talked about, and the authors are careful about it.
It treats perception as a multi-modal, time-evolving alignment problem — exactly the framing that makes modern foundation models good at language and vision. The fact that the same recipe predicts neural activity is a strong signal that these representations are capturing something structurally real about how complex stimuli are processed.
The interesting result isn't "AI predicts the brain." It's that a single alignment objective, applied across modalities, transfers to a substrate it was never trained to imitate.
The connection to agent supervision
This sounds far from agentic mailboxes and schedulers, but here's the thread. If you're designing systems where a human supervises an agent, the bottleneck is human attention — what the person can notice, process, and act on. A predictive model of perception is, eventually, a model of what a supervisor will and won't catch. Interfaces that respect the limits of human attention will beat interfaces that assume infinite vigilance.
We don't use TRIBE in any product. We read it as a marker of where "models of the human" are heading — and that's the half of the loop we think gets under-designed.
The model, weights, and paper are public under a non-commercial license. Worth reading the methods section even if the neuroscience isn't your area; the alignment setup is the transferable idea.