Neural Fields as World Models

Researcher Joshua Nunley has introduced a novel AI architecture in the paper 'Neural Fields as World Models' that challenges conventional approaches to building world models. The core innovation is 'isomorphic world models'—architectures that preserve the spatial structure of sensory input, unlike typical latent-space models that discard this topology. This allows physics prediction to be framed as geometric propagation through a structured neural field, rather than abstract transitions in a compressed latent space.

The technical implementation uses neural fields with motor-gated channels, where activity evolves through local lateral connectivity and motor commands multiplicatively modulate specific neural populations. Three key experiments validate the approach. First, local connectivity proved sufficient to learn ballistic physics, with predictions naturally traversing intermediate spatial locations instead of 'teleporting' between states. Second, and most significantly, policies trained entirely in the model's imagination transferred to real physics environments at nearly twice the rate (a ~100% improvement) compared to standard latent-space world models. Third, the motor-gated channels spontaneously developed body-selective neural encoding purely through visuomotor prediction tasks, without explicit body schema supervision.

This work bridges neuroscience-inspired computation and practical AI. By maintaining spatial isomorphism between the model's internal representations and the sensory world, Nunley's approach creates more intuitive and geometrically grounded predictions. The findings suggest that fundamental cognitive capabilities like intuitive physics and body awareness may emerge from the same underlying principle: spatially structured neural dynamics that mirror the external environment. The paper has been submitted to CogSci 2026, indicating its relevance to both cognitive science and machine learning communities.