Online Generalised Predictive Coding

The key advance is not just adapting generalised filtering to online settings, but doing so through a specific mathematical trick (temporal-scale separation) that avoids the computational trap most real-time systems face: you can't reprocess all historical data every time new observations arrive. The paper shows how to decouple fast state tracking from slower parameter learning, which is the actual engineering constraint that has blocked deployment.

This connects directly to the readmission prediction work from early May, which identified observation window depth as a practical deployment friction point in healthcare ML. That paper benchmarked encoding strategies but didn't address the online inference problem. This generalised filtering paper solves the inverse problem: how to learn and update model structure continuously without batch reprocessing. The temporal decomposition strategy here also echoes the multi-scale approach in MSMixer (also from May 4), which used learned gating to handle oscillations and trends at different timescales. Both papers treat temporal structure as learnable rather than fixed, though in different contexts (forecasting versus filtering).

If robotics or autonomous systems papers cite this work within the next six months as their uncertainty quantification layer, that signals real adoption beyond theory. Alternatively, watch whether any of the major time-series forecasting benchmarks (like those used in MSMixer evaluation) incorporate an online variant of this framework by Q4 2026. If neither happens, the work remains academically sound but practically niche.