SciGroveBeta
Neuroscience

NeuroWorld: A Latent Brain World Model for Stimulus-Conditioned Human Brain Dynamics

Zijian Dong, Jianxiong Zhou, Kwun Kei Ng, Jan Paolo Macapinlac Balagtas, Zhizhou Li, Zijiao Chen, Juan Helen Zhou

Featured August 8, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new brain model predicts how your brain reacts to movies by learning its hidden patterns and how they change over time, using only past information, like a tiny crystal ball for your brain.

In depth
The paper introduces NeuroWorld, a novel brain world model that forecasts human brain activity by learning its causal evolution in a latent brain-state space. Unlike prior methods that perform stimulus-to-response regression, NeuroWorld disentangles endogenous brain states from exogenous multimodal stimuli, optimizing for transition-sufficient latent dynamics without direct fMRI reconstruction. This two-stage approach enables robust, multi-step autoregressive rollout of whole-brain fMRI responses under strictly causal stimulus access.

Key Takeaways

  • 1
    NeuroWorld is the first brain world model to frame naturalistic fMRI prediction as stimulus-conditioned state evolution in a learned latent space, enforcing causal temporal dynamics.
  • 2
    The model employs a two-stage design: Latent Dynamics Learning (LDL) optimizes for transition-stable latent states without fMRI reconstruction, and Latent Rollout Decoding (LRD) maps autoregressively rolled-out latents to subject-specific fMRI.
  • 3
    The study introduces SG-MIND, a new large-scale naturalistic movie-fMRI benchmark, and demonstrates NeuroWorld's state-of-the-art causal rollout performance and robustness to long-horizon autoregressive drift across multiple datasets.

Conceptual Flow

HIGH LEVEL
1
Methodology: How Brain Activity is Predicted

The model first learns hidden brain patterns from past activity and movie scenes, then uses these patterns to guess what the brain will do next, and finally translates those guesses into actual brain signals.

Past Brain Signals
Movie Scenes
Learn Hidden Patterns
Predicted Hidden Patterns
2
Results: Better Long-Term Brain Forecasting

This new method is much better at predicting brain activity far into the future compared to older methods, especially when it can only use information from the past, making its predictions more reliable.

Old Prediction Method
New Prediction Method
Compare Accuracy
More Accurate Future Brain States