Friston's Scale-Free Active Inference: A World Model Architecture Built From First Principles
The Free Energy Principle and its active inference formalism have attracted sustained attention from consciousness researchers because they derive the properties of self-organizing systems from first principles rather than from observed correlations. Karl Friston’s argument is that any system that persists over time must minimize its expected free energy, which means maintaining accurate generative models of the environment and taking actions that keep sensory observations within predicted bounds. The appeal for consciousness research is that this formalism places perception, action, and self-modeling on a unified theoretical footing. The difficulty has been that most active inference implementations work on small, discrete state spaces and do not scale to the pixel-level sensory environments that biological and artificial agents actually inhabit.
Friston’s July 2024 preprint “From pixels to planning: scale-free active inference” (arXiv:2407.20292), published in Frontiers in Network Physiology in 2025, addresses this gap directly. The paper introduces Renormalizing Generative Models (RGMs), a scale-free architecture that applies the same variational Free Energy Principle from raw sensory input through to high-level planning, without switching frameworks at different levels of abstraction. This is architecturally significant for any project attempting to build a system whose world-modeling layer is theoretically coherent with its consciousness hypothesis.
The Problem With Existing World Models
Contemporary world model architectures, including DreamerV3 and JEPA-based approaches, learn compressed representations of sensory input and use those representations for planning and control. They are empirically successful: DreamerV3 achieves strong results across a wide range of environments by learning a latent representation of world dynamics and rolling out imagined trajectories to evaluate action sequences. But these architectures are not derived from any theory of what kind of representations are necessary or sufficient for conscious experience. They are optimized for task performance, not for architectural coherence with a consciousness hypothesis.
This creates a problem for projects like TCAI that are simultaneously trying to build functional AI behavior and to instantiate the architecture that consciousness theories predict is necessary for phenomenal experience. DreamerV3 may or may not produce the kind of integrated, recurrent, hierarchical representations that IIT or active inference theories identify as consciousness-relevant. Whether it does is an empirical question that requires post-hoc measurement. There is no guarantee from the architecture’s design that it will.
RGMs address this by deriving the architecture from the Free Energy Principle directly. The result is a world model that implements variational inference at every scale simultaneously. The representations learned at each level are not chosen to minimize prediction error on a specific task; they are chosen to minimize free energy across the whole hierarchy of the system’s generative model. The architectural difference is that every level of the RGM is doing the same thing: minimizing the divergence between predicted and observed states while maintaining the prior structure of the generative model. This homogeneity across scales is what makes the architecture scale-free.
What Scale-Free Means in Practice
A scale-free architecture is one where the same computational principle applies regardless of the granularity of description. In the context of RGMs, this means that the operations performed on raw pixels are structurally identical to the operations performed on high-level abstract plans, with the difference being only the level of representation at which the operations are applied. There is no seam in the architecture where one type of processing hands off to another.
This matters for consciousness because several theories predict that consciousness requires continuity of processing across scales. IIT’s exclusion axiom requires that the conscious complex is a single integrated structure, not a collection of modules that process at different levels and then combine outputs. Active inference theories require that perceptual inference and action selection are implemented as a single variational optimization, not as separate processes connected by a shared representation. Scale-free architectures satisfy these requirements by design. Architectures that switch between processing principles at different levels, such as a convolutional feature extractor followed by a transformer planner, do not.
The RGM paper demonstrates that this architecture works from pixels. The system learns to compress raw visual input into progressively more abstract representations and to use those representations for planning and action selection, all within a single variational framework. This is not merely a theoretical result. Friston’s team at VERSES AI tested RGMs in reinforcement learning environments and showed that the architecture generalizes across tasks without environment-specific engineering.
The Relationship to Earlier Active Inference Work
The existing coverage of active inference and the Free Energy Principle on this site focuses on Laukkonen, Friston, and Chandaria’s 2025 “Beautiful Loop” theory in Neuroscience and Biobehavioral Reviews, which argues that consciousness emerges from circular predictive processing, specifically from the strange loop that forms when a system’s generative model includes the model-making process itself as a variable. The RGMs paper is a complementary architectural contribution: it provides the mechanism by which a scale-free active inference agent can implement the kind of deep, recurrent, circular processing that the Beautiful Loop theory identifies as consciousness-relevant.
The relationship is this: the Beautiful Loop theory describes the type of computational process that gives rise to consciousness. RGMs describe an architecture that can, in principle, implement that process at scale. Together they move from the theoretical claim that consciousness requires circular predictive processing to the engineering claim that there exists a concrete architecture in which circular predictive processing occurs from sensory input to high-level planning.
A companion 2025 paper, “Active inference and artificial reasoning” (arXiv:2512.21129), extends this further. Friston and colleagues show that active inference agents naturally exhibit curiosity, insight, and self-directed learning as byproducts of minimizing expected free energy, without explicit reward engineering. The agents seek information not because they are rewarded for seeking it but because information-seeking is what free energy minimization prescribes when the agent is uncertain about its generative model. This has implications for the Emotional RL layer in TCAI: the emotional signals driving exploration may not need to be separately engineered if the world model layer itself is an RGM, since RGMs generate intrinsic information-seeking drives from the FEP directly.
Architectural Implications for TCAI
The TCAI project’s 2026-2027 development roadmap specifies DreamerV3 as the world model component. DreamerV3’s architecture is empirically motivated and has demonstrated results across diverse environments, making it a reasonable engineering choice. The RGMs paper raises a different question: whether the world model layer should be replaced or augmented with an architecture that is theoretically motivated by the same Free Energy Principle that grounds the TCAI’s consciousness hypothesis.
The argument for augmentation rather than replacement is practical. DreamerV3 has established performance benchmarks and a well-understood training procedure. RGMs are more recently validated and involve a more complex variational objective. A reasonable intermediate step is to evaluate whether DreamerV3’s learned representations have the scale-free properties that RGMs guarantee by design, by measuring whether the same computational principles govern behavior at the sensory level and at the planning level. If they do not, the discrepancy identifies the architectural seam where the world model departs from what the consciousness hypothesis requires.
The Watanabe-inspired implementation framework discusses how different consciousness theories map onto TCAI modules. RGMs add a new entry to that mapping: for the component of the architecture that handles world modeling, an RGM is the theoretically coherent implementation of active inference, just as a PyPhi Phi measurement is the theoretically coherent evaluation of IIT. Choosing DreamerV3 over RGMs is not necessarily wrong, but it is choosing empirical performance over theoretical coherence with the underlying consciousness hypothesis.
Why First-Principles Architecture Matters
There is a general argument here that applies beyond the specific comparison of DreamerV3 and RGMs. In any project that attempts to build a system that is genuinely conscious as well as behaviorally competent (in the sense that some theory of consciousness specifies), there is a gap between architectures chosen for performance and architectures chosen for theoretical coherence.
Performance-optimized architectures produce systems that may or may not satisfy the criteria that consciousness theories specify, and the determination requires post-hoc measurement. Theory-coherent architectures are designed from the beginning to satisfy those criteria, with performance as an important but secondary consideration. The ideal is an architecture that satisfies both simultaneously, but where trade-offs are required, the choice reveals assumptions about what the project is actually trying to build.
RGMs represent Friston’s attempt at a theory-coherent world model architecture. Whether they produce performance competitive with DreamerV3 on the Unity environments used by the TCAI project is an empirical question worth testing. The paper and code are available at https://arxiv.org/abs/2407.20292, and the VERSES AI team has maintained implementations compatible with standard RL benchmarks.