NeuroWorld: First brain world model forecasts fMRI during naturalistic movies

NeuroWorld, developed by Zijian Dong and colleagues at the National University of Singapore, is a novel brain encoding framework that reframes fMRI prediction as a causal world-model problem. Unlike traditional stimulus-to-response regressors that leak future information, NeuroWorld learns a latent brain-state space where neural dynamics evolve under continuous sensory input. The architecture separates endogenous states from exogenous multimodal stimuli in two stages: Latent Dynamics Learning (LDL) jointly learns a transition-sufficient representation and causal dynamics via next-latent prediction—without reconstructing raw fMRI—while Latent Rollout Decoding (LRD) freezes LDL and autoregressively decodes latent rollouts into whole-brain maps.

The model was validated on three naturalistic movie-fMRI benchmarks covering 30 participants, including the newly collected SG-MIND dataset (20 participants, 8,519 paired stimulus-response clips, 140.7 person-hours of viewing). NeuroWorld achieves state-of-the-art multi-step rollout performance under strictly causal stimulus access, with greater robustness to long-horizon autoregressive drift than prior methods. Interpretability analyses revealed structured functional organization in the learned dynamics, suggesting the latent space captures meaningful brain states. This work positions world modeling as a principled framework for causal forecasting of human brain activity, with potential applications in cognitive screening, brain-computer interfaces, and neurofeedback systems.