Meta Unveils AI Model That Predicts Human Brain Activity
Meta has introduced TRIBE v2, a multimodal AI model designed to predict how the human brain responds to visual, auditory, and language stimuli. As a result, the system could accelerate neuroscience research while reducing reliance on costly brain-scanning experiments.
The TRImodal Brain Encoder v2 builds on an architecture that previously won first place at the Algonauts 2025 brain encoding competition, outperforming more than 260 teams. Moreover, researchers trained the model on fMRI recordings from over 700 volunteers who watched movies, listened to podcasts, and read text in multiple languages during scanning sessions.
Compared to its predecessor, TRIBE v2 represents a major leap in scale. While the original version used data from four subjects and around 80 hours of recordings, the new model incorporates more than 1,000 hours of fMRI data from 720 participants. Consequently, it predicts responses across roughly 70,000 brain voxels, increasing spatial resolution by about 70 times.
In addition, the model uses a three-stage pipeline. First, it extracts features through separate pretrained encoders for video, audio, and text. Then, it combines these signals using a transformer-based integration module. Finally, it maps the output onto the cortical surface to generate predictions.
Toward In-Silico Neuroscience
One of the model’s defining capabilities is “zero-shot prediction,” which allows it to forecast brain responses for new individuals, unseen languages, and entirely new tasks without retraining. Because of this, researchers reported two- to three-fold accuracy improvements, especially for subjects not included in training data.
This capability enables what researchers call “in-silico experimentation.” For instance, when given a sentence, TRIBE v2 can predict fMRI activity across known language-processing regions and reproduce patterns typically observed in human studies. Therefore, scientists may test hypotheses computationally before conducting expensive real-world experiments.
At the same time, traditional fMRI sessions can cost thousands of dollars and require institutional review board approval. By contrast, computational simulations offer a faster and more flexible alternative, which could significantly streamline early-stage research.
Open Access and Future Potential
Meta has released the model weights, codebase, research paper, and an interactive demo under a CC-BY-NC-4.0 license. Through this approach, the company aims to support broader research efforts and encourage innovation in both neuroscience and AI development.
In its official announcement, the company framed the release as a step toward applying brain insights to build better AI and speeding up breakthroughs in neurological disease diagnosis and treatment.
However, brain encoding models still explain only a limited portion of the variance in fMRI data, as several researchers have noted. Even so, the availability of a large dataset and open tools provides a stronger foundation for further testing. Ultimately, the model’s real-world impact will depend on how independent scientists evaluate and apply it in their own research.
