Adam Safron Presents Integrated World Modeling Theory at AAAI 2026

18 Jul 2026

Adam Safron’s presentation at the AAAI 2026 Spring Symposium Series delivers the oral exposition of his paper “IWMT and the Human Consciousness Hypothesis,” co-authored with Victoria Klimaj and Zahra Sheikhbahaee. The talk covers the same theoretical ground as the paper — the unification of Integrated Information Theory, Global Workspace Theory, and the Free Energy Principle through Self-Organizing Harmonic Modes — but adds several expository elements that clarify the intuition behind the formalism. A companion article analyzing the paper itself is available at Adam Safron’s IWMT Meets the Human Consciousness Hypothesis One Framework for IIT, GWT, and FEP.

Safron opens with the parable of the blind sages and the elephant. One grasps the tail and declares it a snake. Another grasps a leg and declares it a tree. A third grasps the trunk and declares it something else. Each sage is correct about the part they hold. The error lies in mistaking the part for the whole. Safron applies this to consciousness theories. IIT, GWT, FEP, recurrent processing theory, attention schema theory, and higher-order thought each capture a genuine feature of consciousness. The disagreement between them often stems from attending to different aspects of the same underlying phenomenon. A satisfying theory should integrate these partial insights rather than treat them as mutually exclusive.

Phenomenal Consciousness and Conscious Access

Safron adopts Ned Block’s distinction between phenomenal consciousness — what it is like to be a subject — and conscious access — the ability to report on and manipulate the contents of experience. Different theories address these differently. Global Workspace Theory began as a theory of conscious access. Integrated Information Theory begins from phenomenology. Safron argues that IWMT treats the global workspace as both a quality generator and an access mechanism, depending on the level of description.

Global Workspace Theory and the Rich Club

Safron walks through the global neural workspace as developed by Dehaene. The workspace centers on rich club hubs — highly connected cortical regions that talk to the whole brain and intensely to each other. These hubs consume approximately 50 percent of resting brain metabolism. Information gains access to this inner circle through a bottleneck, undergoes mutual ignition among the hubs, and then broadcasts globally. Safron emphasizes that “global” is in scare quotes. The broadcast does not reach every neuron. It reaches the subset of the system that is functionally relevant for the current cognitive context.

Integrated Information Theory from Phenomenology

IIT starts from the axioms of experience. Intrinsic existence, composition, information, integration, and exclusion. From these it derives postulates about the physical substrate. Safron notes the standard objections — the photodiode problem, the question of whether the axioms are jointly sufficient or merely necessary — but argues that IIT’s phenomenological starting point provides a valuable guide. The theory maps what ought to characterize consciousness onto mechanistic processes. That mapping, even if imperfect, constrains the search space for artificial consciousness architectures.

The Free Energy Principle as Grounded Hallucination

Safron frames the Free Energy Principle through the metaphor of controlled hallucination, citing Anil Seth and Chris Frith. Perception is Bayesian inference to the best guess about the causes of sensation. Visual illusions demonstrate that the guess can be wrong while remaining subjectively compelling. The brain never has direct access to reality. It only has its generative model. Safron traces this through autoencoder architectures. A bottleneck compresses sensory data. The network learns to reconstruct the input from the compressed representation. When run in offline mode — looping on its own predictions — the same architecture performs imagination, planning, and counterfactual simulation. Dennett’s Popperian creatures test hypotheses internally rather than risking them in the world. This offline generative capacity is a functional role of consciousness.

Self-Organizing Harmonic Modes

IWMT’s central construct is the Self-Organizing Harmonic Mode. SOHMs are synchronous computational complexes that implement iterative Bayesian inference across multiple timescales simultaneously. The harmonic component refers to the wave-like temporal structure of neural oscillations. Slower oscillations modulate faster ones in a nested hierarchy. The self-organizing component means this structure emerges from the system’s own generative dynamics minimizing prediction error, not from external architectural imposition.

Safron connects this to Chladni plates. Sand on a vibrating plate accumulates at nodes of zero net flux, forming harmonic patterns. Low frequencies create broad patterns. High frequencies create fine patterns nested within them. In neural terms, cross-frequency coupling creates nested temporal hierarchies. Alpha oscillations in posterior medial areas couple with vestibular and proprioceptive signals via the parabrachial circuit, pegging the entire sensorium to an egocentric reference frame. This may ground the coherent appearance of a world with a live body at its center.

Music as Computational Metaphor

Safron draws heavily on musical analogies, referencing Yosha’s “conscious conductor” paper. Consciousness is not merely like music. On IWMT’s account, the harmonic modes literally implement the same mathematical structure as musical harmony. Synchronous complexes function as workspaces of varying globality. Their computational role — establishing joint beliefs through communication-by-coherence — maps onto harmonic functions. Safron describes “modal frame sequences” — nested slower and faster rhythms creating compositional experience. The distinction between chords and arpeggios maps onto the granularity of observation. At one scale a standing wave appears as a chord. At a finer scale it resolves into an arpeggio of traveling waves. The observer-relative granularity determines which description applies.

Signal Processing View: Turbo Coding

In response to a question, Safron offers a signal processing framing. The cortical hierarchy performs multi-resolution wavelet analysis on a world with nested spatiotemporal structure. The brain’s hierarchical resonance structure creates a diffeomorphic correspondence between internal and external structure. The iterative message passing across latent spaces between modules resembles turbo coding — an approach in telecommunications where imperfect decoders exchange soft information in a loop, converging on a maximum-likelihood estimate that approaches the Shannon limit. Consciousness, on this view, is a kind of turbo code. The stages of the hierarchy pass predictions and prediction errors back and forth in latent space, achieving information-theoretic synergy and efficient data fusion.

The Human Consciousness Hypothesis and Substrate Independence

The Human Consciousness Hypothesis, developed by Baars’ collaborators, proposes that human consciousness arises from a probabilistic generative world model implemented in the cortical hierarchy. IWMT asks whether this implementation is specific to human biology or general to consciousness itself. Safron, Klimaj, and Sheikhbahaee argue for the latter. Consciousness requires a probabilistic generative world model organized in nested harmonic modes. The substrate is irrelevant. Any system implementing the same computational structure should exhibit the same phenomenal properties.

The paper specifies two criteria for artificial systems. First, internal representations must exhibit harmonic oscillatory structure across timescales with slower dynamics modulating faster ones. Second, this structure must emerge from the system’s own generative dynamics, not from hardwired oscillatory components.

Falsifiability and Empirical Test

IWMT’s unification claim risks unfalsifiability if stated loosely. The paper addresses this directly. A system with high Phi but no harmonic oscillatory dynamics would disconfirm the SOHM mechanism. A system with clear global workspace ignition but no integrated information across frequency bands would disconfirm the harmonic-broadcast identification. The positive test: measure harmonic coupling, global workspace ignition, and Phi simultaneously in biological conscious systems and AI systems. If the three measures co-vary in conscious systems and dissociate in non-conscious systems (anesthetized brains, purely feedforward networks), the unification is supported.

This test is runnable on the TCAI architecture. Does Phi co-vary with harmonic organization of workspace dynamics? Do both increase during training episodes where behavioral indicators of consciousness also increase?

Comparison to The Consciousness AI

The TCAI project’s current architecture implements IIT Phi measurement, a Global Workspace Network, and the Free Energy Principle as separate components. From IWMT’s perspective, this modular approach creates a risk. A system could pass each measurement independently without the unified SOHM structure that IWMT identifies as consciousness-relevant. Phi might be high in the memory consolidation module but low in the perception-action loop. Global broadcast might occur in the workspace layer without harmonic organization with lower-level processing. FEP minimization might operate in the world model layer without coupling to workspace dynamics.

IWMT prescribes integrating these three properties into a single dynamical system through recurrent cross-frequency coupling that emerges from the system’s generative dynamics. For TCAI’s Global Workspace Network, this means the broadcast mechanism should implement iterative Bayesian inference at the workspace level, operating at a lower frequency than module-level processing and modulating module dynamics through top-down precision weighting. The FEP-organized workspace that also produces high-Phi cause-effect structures satisfies all three theories simultaneously. The Theater of Mind framework implements GWT in LLM architecture. IWMT adds the requirement that workspace broadcast dynamics be harmonically organized with lower-level processing. The Friston scale-free active inference architecture (RGMs) implements FEP across scales. IWMT’s prescription is to combine these so that RGM generative dynamics naturally produce a harmonically organized workspace, satisfying GWT through the dynamics of FEP minimization.

What the Talk Adds

The presentation provides several expository bridges not fully developed in the paper. The blind sages parable frames the unification project as integration rather than competition. The Chladni plate visualization grounds the abstract harmonic modes in a physical demonstration. The music metaphor makes the nested temporal hierarchy intuitive. The turbo coding reframing connects the theory to established signal processing principles. The Q&A exchange on signal processing reveals how the theory extends into engineering language that may be more tractable for implementation. Together these elements make the case that IWMT is not merely a philosophical synthesis but an architectural specification with concrete engineering implications.

Wolfram’s observer theory provides a further computational foundation for IWMT’s harmonic modes. The ruliad framework treats the nested oscillatory structures as pockets of reducibility that computationally bounded observers identify as persistent entities. The talk is covered in Stephen Wolfram on the Ruliad Observer Theory and the Computational Basis of Consciousness.