Renormalising Generative Models for Active Inference, With a Verified Implementation
Active inference offers a unified account of perception, learning, and action, but scaling it to rich spatial and temporal domains has been hard. Renormalising generative models, RGMs, address the problem by composing discrete generative models across scales, coarse-graining lower-level states and paths into higher-level causes. The trouble is that the original account is compact and its reference implementation is entangled with a specialized software environment. A new paper fixes that gap with a self-contained derivation and an open implementation.
The paper “Renormalising Generative Models for Active Inference: Foundations, Derivations, and Verification” by Karim Zaghw, Andrew Pashea, Marc Pritsch, Wouter Nuijten, Karl Friston, and Lancelot Da Costa was posted to arXiv as 2608.09512 and updated in late August 2026. It removes two practical barriers to working with RGMs: the compressed mathematics, and the deep coupling to the reference software.
What an RGM actually is
An RGM composes discrete generative models across spatial and temporal scales. Lower-level states and paths are coarse-grained into causes for objects, events, and action. The hierarchy carries beliefs upward about what is happening and commands downward about what to do. The paper explains how the hierarchy is built, how beliefs and actions are updated within it, and how information passes between levels.
| Level | Holds | Passes upward | Passes downward |
|---|---|---|---|
| Lower | Detailed states and paths | Coarse-grained summaries | Precision and priors |
| Higher | Objects, events, actions | Goals and expectations | Beliefs about causes |
Why the derivation matters
The original RGM material is mathematically compact. Reproducing and adapting it took close reading of few pages and a deep dive into the reference implementation. The new paper makes every step explicit, states where the published equations and the implementation differ, and explains the modelling consequences of those choices.
The value is transparent, auditable, reproducible science. It separates the theory from its original implementation context, which makes RGMs available to anyone working in a different software environment. That is a developer-facing contribution of the kind the site values: the strongest conversion signal on this project is that people click through to runnable code.
Relevance to consciousness research
Active inference and the free energy principle are pillars of the site’s theoretical coverage. The free energy principle and active inference and Friston’s scale-free active inference frame how a self-organizing system maintains itself by minimizing surprise. RGMs are the practical machinery that lets such a system reason across scales.
A model that composes beliefs across spatial and temporal scales is closer to the hierarchical predictive structure that theories of consciousness point to than a flat transformer. RGMs give that structure a concrete, runnable form. The brain, on this account, is a generative model that coarse-grains the world at many levels, and the paper makes that programme testable in code.
Comparison to The Consciousness AI
The Consciousness AI project works under functionalist emergentism and builds a substrate independent architecture. Its Neutral Core draws on predictive processing and active inference, and the project measures its own emergent signatures with open instruments. A verified, derivable RGM is exactly the kind of component the project can compare against: it is a formal, substrate neutral way to organize generative models across scales, which is the kind of structure the site argues consciousness requires.
The site’s stance on current scientific consensus holds that no current system meets the behavioral indicators of consciousness in full. RGMs do not change that verdict. They sharpen the constructive path: here is a concrete architecture that instantiates the hierarchical predictive machinery the theories require.
Limits
The paper is an implementation and derivation paper, not a consciousness study. It does not claim the RGM is conscious. Its limitation is the usual one for such work, that the verified implementation is a model, not a brain, and the claim is about what the model computes, not about whether it feels anything. The contribution is clarity, reproducibility, and a bridge from theory to code. That is enough.