ICLR 2025 · 2025

FACTS: A Factored State-Space Framework for World Modelling

Li Nanbo, Firas Laakom, Yucheng Xu, Wenyi Wang, Jürgen Schmidhuber

Paper
Code

A world model needs to remember several evolving entities and their interactions. FACTS replaces a single undifferentiated recurrent state with a factored, graph-structured memory, bringing structured representations together with efficient state-space sequence modelling.

Factored memory over time

FACTS diagram showing differently ordered input entities routed to persistent colored memory streams over successive time steps. — Learned routing connects observations to factored memory streams, allowing the representation to track factors as their ordering in the inputs changes. View full size · Source

Routing, structure, and propagation

A routing mechanism learns permutable memory representations so that the model can handle changes in input ordering. Selective state-space propagation updates these representations over time, while the graph structure supports relationships between factors. The framework also supports parallel computation for high-dimensional sequences.

The paper evaluates this approach on multivariate time-series forecasting, object-centric world modelling, and spatial-temporal graph prediction, exploring how a shared structured-memory framework applies across these settings.