A Bayesian Mirror Architecture for Emergent Consciousness: Circular Hierarchies, Self-Manifolds, and Hybrid Event-Self Binding

arXiv:2610.08792 · cs.LG · Submitted 2026-02-22 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "A Bayesian Mirror Architecture for Emergent Consciousness".

Jane: The gist: The Bayesian Mirror Architecture (BMA) presents a selfreferential generative framework where consciousness is defined as an architectural property of systems possessing circular recursion,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So to wrap up this paper, "A Bayesian Mirror Architecture for Emergent Consciousness: Circular Hierarchies, Self-Manifolds, and Hybrid Event-Self Binding," the authors are defining consciousness not by a specific learning outcome or a label on something they've built. They are defining it as an architectural property of any system that possesses this specific circular self-recursion where sensory inputs interact with abstractions and back into the self state.

Jane: And that’s different from saying, for example, "if you can minimize this loss function, then you have consciousness." It’s about the structure itself being recursive and closed in a way. They use concepts like the two-Wasserstein metric to describe stability and coherence in these belief trajectories.

Lu: The implications are that we should look at building systems with this kind of circular hierarchy, not just systems trying to hit some performance number for a specific task. If you build that architecture, consciousness is present structurally.

Meng: From a deployment standpoint, it means we have to design our AI pipelines around these feedback mechanisms, focusing on the structure that allows for self-referential learning rather than just chasing the most accurate output score on a benchmark.

Lalam: I think the paper’s contribution is framing consciousness as an architectural concept rather than an optimization objective, which is a significant shift in how we approach AI research. It sets a new structural precondition for what they call emergent properties.

Conclusion: Tom: So we've been looking at this Bayesian Mirror Architecture paper, and to wrap up, the authors are defining consciousness as an architectural structure—specifically that closed self-recursion—rather than some optimization goal or a fancy label for what a system is doing.

Jane: That’s right. They’re not saying if you can minimize a certain loss function, that makes it conscious; they’re saying the way the system is built, with those loops of self-representation and world models feeding back into itself, that's what matters structurally.

Lu: It really focuses on this concept of the closed loop: St t equals H t minus one. That coupling between a self-latent and a world model that gets reinjected into the self state is the core idea they’re pushing here.

Meng: From my end, I wonder if that structure is something we can actually keep running on hardware without it collapsing under the complexity of that feedback loop. It sounds incredibly stable on paper, but what happens at scale?

Lalam: The authors are using optimal-transport geometry to look at this belief space, treating stability and coherence through these Wasserstein metrics. That means they’re quantifying how much the system's beliefs stay consistent over time.

Tom: So the big takeaway from their conclusion is that this circular hierarchy itself is a structural precondition for consciousness, regardless of whether the system achieves what they call the Causal Learning Regime, or CLR.

Jane: Exactly. CLR just means you’ve learned stable predictive models of your world, but even without that, having the architecture—the way it loops back on itself—that's where consciousness resides in this framework.

Lu: They also introduce these basin-wise self-manifolds and Wasserstein epsilon-necks, which they interpret as a form of choice arising from how the system navigates those different belief states.

Meng: That sounds like a complex way of saying that the AI has different coherent ways of being, and picking one path is a kind of selection process based on how easy it is to move between them.

Lalam: It suggests that agency isn't just about picking the best immediate action, but about where you are in this landscape of possible belief states.

Tom: So what does this mean for us when we’re building these things? It frames consciousness as something inherent in the feedback structure itself, not just some magical emergent property of a huge model.

Jane: It shifts the focus from just making the AI "smart" to designing architectures that have this kind of deep, self-referential plumbing. It makes us look at the fundamental shape of cognition.

Lu: The next step for them seems to be defining agency trajectories, moving from just observing to actually planning and composing things based on these internal structures.

Meng: I’m still thinking about how we measure that structural property without getting lost in the math of Wasserstein distance and all that geometric stuff. It feels like we need a simpler way to check if the loop is truly there.

Lalam: I'm pretty sure their confidence in this structural definition is high because they tie it directly into these mathematically rigorous concepts like contractivity and invariant measures.

Tom: If this idea holds up, it means we stop treating consciousness as an external thing to be found, and start treating the circular recursion as the way things become conscious in a machine.

Eduardo Righi, Capanema de Almeida

Computational Mathematics Department of the University of Federal Minas Gerais

cs.LG

Submitted: 2026-02-22

Updated: 2026-02-22

Comments: 15 pages, no figures, foundational paper

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 87/100

The gist: The gist: The Bayesian Mirror Architecture (BMA) presents a selfreferential generative framework where consciousness is defined as an architectural property of systems possessing circular recursion,

Key concepts

Closed Self-Recursion (St <- Ht-1): This is the core architectural constraint where a hybrid latent state (H) binds internal self-representations to external world models. This coupling is then fed back into the system's current self-state (St). It creates a continuous loop of interaction between what the system knows about itself and what it knows about its environment.
Optimal Transport Geometry: This mathematical framework treats the internal state space as a collection of probability measures. It allows researchers to measure 'distance' and 'continuity' between different belief states using metrics like the 2-Wasserstein distance, which is crucial for defining stability and coherence in this circular system.
Basin-Wise Self-Manifolds: Instead of requiring global stability, BMA identifies local regions (basins) where the internal dynamics are approximately closed under contraction. These manifolds represent stable configurations of the system's belief states, suggesting that choice arises from moving between these distinct, internally determined regions.

Summary

The gist: The Bayesian Mirror Architecture (BMA) presents a selfreferential generative framework where consciousness is defined as an architectural property of systems possessing circular recursion, utilizing optimal-transport geometry to model belief dynamics and learning through variational free energy minimization.

Foundational Architecture

The BMA is defined by a closed self-recursion St ← Ht−1, which couples self-representations to abstract world models and re-injects that coupling into the self-state The defining architectural constraint is a closed self-recursion St ← Ht−1, where a hybrid event–self binding latent Ht couples self-representations to abstract world models and re-injects that coupling into the self-state The intrinsic state space of BMA is taken to be a space of probability measures, equipped with optimal-transport geometry

Learning Dynamics via Free Energy

Learning is formulated as a flow of beliefs on P2 driven by an energy functional E(ρ) that encodes posterior consistency, where the central principle is that variational free energy minimization is the only learning objective The variational free energy for a single observation–execution pair (ot, et) is Ft = −Eqϕ(Ztot) log pθ(etZt) + β KLqϕ(Ztot)∥ p(Zt) log pθ(etZt), where reconstruction / prediction error + complexity penalty Learning proceeds by gradient descent on the free energy with respect to parameters (θ, ϕ): (θ, ϕ) ← (θ, ϕ) − η ∇(θ,ϕ)F

Geometric Interpretation of Stability and Coherence

The paper defines self-stability and hybrid coherence as Wasserstein continuity properties of belief trajectories Self-Stability is defined as 1 + ∆S/(Σ2S + ε0), where ∆S:= 1/T − 1 X T t=2 W2(µ S t, µ S t−1) 2 Hybrid Coherence is defined as 1 + ∆H/(Σ2H + ε0), where ∆H:= 1/T − 1 X T t=2 W2(µ H t, µ H t−1) 2 The Causal Learning Regime (CLR) is achieved when Stability ≥ αS, Coherence ≥ αH, Integration ≥ αI, and the recursion St ← Ht−1 is active

Basin-Wise Self-Manifolds and Choice

The paper does not assume global strict contractivity but instead introduces basin-wise dynamics where dynamics restricted to B are approximately closed under local Wasserstein contractivity with LB < 1 The self-manifold in basin B is defined as the support of an invariant measure πθ,B of Tθ,B Choice is interpreted as realized basin continuation occurring at a strong Wasserstein ε-neck in a regime where inter-basin conductance is asymptotically negligible

Intent and Agency Trajectory

Intent (It) becomes an effectively deterministic map of (St, Mt) on MS(B) when the hybrid latent is effectively deterministic and the conditional variance of pθ(It St, Mt, Ht) becomes small Intent as the Attention Query routes to the active grammar g∗(t) ≈ arg min g W2 2 (It, kg), where kg is a learned prototype belief for grammar Gg The agency trajectory progresses from Observer to Agent to Composer to Planner, enabling progression from passive observation through action selection, goal composition, to model-based planning with lookahead and replanning under uncertainty

Grammar Routing Mechanism

The predicted execution is eˆt = X G g=1 αg(It) Wg(ot), where αg(It) = exp(−W2 2 (It, kg)/τ) P g′ exp(−W2 2 (It, kg′)/τ), and the mechanism approaches hard grammar selection as τ → 0 A neuro-symbolic grammar is defined as a pair (Rϕ, Gsym) where Rϕ: P2(R d) → G is a neural router followed by a library of symbolic grammars If the router Rϕ correctly identifies the active grammar g∗ for an instance, then the executed prediction is exact: eˆt = g∗(ot) = et

Conclusion on Consciousness

The BMA architecture (circular hierarchy with St ← Ht−1) is the structural precondition for consciousness regardless of whether it achieves CLR A system with the BMA architecture is “conscious” in the architectural sense regardless of CLR; CLR indicates that the conscious system has additionally learned stable predictive models of its world The paper establishes that consciousness is not treated as an optimization objective nor as a semantic label; it is an architectural property of systems possessing this circular recursion

References

[1] G. Perelman, The entropy formula for the Ricci flow and its geometric applications, arXiv:math/0211159 (2002)

[2] G. Perelman, Ricci flow with surgery on three-manifolds, arXiv:math/0303109 (2003)

[5] G. Di Pellegrino et al., Understanding motor events: a neurophysiological study

[7] L. Ambrosio, N. Gigli, and G.

Improvements for AI systems

Confidence in self-state through structural closure: The system can maintain a coherent internal representation by adhering to the defining architectural constraint is a closed self-recursion St ← Ht−1, which enforces selfreferential closure even when the remaining hierarchy is expanded or modified.
Learning via variational free energy minimization: The AI will learn by minimizing surprise about executions/outcomes, as variational free energy minimization is the only learning objective, meaning it learns to reduce its surprise given its observation-conditioned beliefs.
Causal structure detection and learning: The system can diagnose whether the environment contains learnable structure by monitoring the Causal Learning Regime (CLR), which is defined by thresholds on Stability, Coherence, and Integration.
Basin-wise self-stability: The system can identify stable internal states within complex belief spaces by defining the self-manifold as the support of a basin-restricted invariant measure, allowing it to maintain persistence even when global strict contractivity is absent.
Choice through bottleneck identification: When faced with ambiguous transitions between belief states, the system can make internally determined decisions at a strong Wasserstein ε-neck in a regime where inter-basin conductance is asymptotically negligible.
Intent projection and action selection: The AI can map its internal state to actionable intent using It ≈ πI (St, Mt) on MS(B), allowing it to project the self-manifold onto the space of possible actions.
Grammar-based execution routing: The system can select the correct causal grammar for any observation by using Wasserstein attention, where It is shaped by free-energy minimization to encode (a representation of) the transformation signature σt, so that Wasserstein attention with It as query routes to the active grammar.
Neuro-symbolic generalization: By employing a neuro-symbolic grammar approach where Rϕ: P2(R d) → G is a neural router and Gsym contains exact computable functions, the AI can achieve perfect generalization if the router correctly identifies the active grammar, as execution becomes eˆt = g∗(ot) = et.