Consciousness as Entropy Reduction

• CER is a model where consciousness emerges as a determined scenario selected via entropy reduction from a probabilistic ensemble.
• It employs a weighted entropy functional and gradient descent (S2C) to integrate multi-channel feature maps into a unified conscious experience.
• The framework connects with Global Workspace Theory and IIT, explaining internal simulation and self-referential feedback in conscious processing.

Consciousness as Entropy Reduction (CER) is a model of consciousness in which subconscious contents are represented as probability distributions over possible “scenarios,” and a conscious experience arises when entropy-reducing gradient descent selects a single determined scenario from that space. In this formulation, the feature map is not itself consciousness, but “the input scenario into a world of possible subconscious scenarios from which the conscious scenario (i.e., conscious experience) is chosen.” CER was introduced as a logically grounded model inspired by Global Workspace Theory (GWT), Integrated Information Theory (IIT), and an earlier “Feature Map” approach in psychology, and it presents conscious processing as an internal simulation of the outside world in which solving problems internally is more economical than acting them out directly in the environment (Chen et al., 7 Oct 2025).

1. Scenario ontology and the definition of conscious content

CER starts from the claim that the contents of consciousness and subconsciousness are best treated as scenarios: vectors of patterns or features distributed across channels or feature locations. A single channel carries a probability distribution over patterns, $x:P\to[0,1]$, with $\sum_{p\in P} x(p)=1$. A multi-channel scenario is then a distribution over $P^n$, so that a full scenario specifies one pattern per channel. A distribution is “determined” when it is a point mass, and CER identifies conscious experience with precisely such determined scenarios (Chen et al., 7 Oct 2025).

The model draws a sharp distinction between subconscious and conscious representation. The subconscious is a probability distribution over many possible world states, and at any moment it is assumed to be fully unintegrated, meaning that the joint scenario distribution factorizes into the product of single-channel distributions. Consciousness is the opposite limit: a single, fully specified scenario. In the short-version formulation, conscious processing is therefore identified with entropy reduction from a diffuse, unintegrated probabilistic scenario to a “zero-entropy, deterministic scenario” (Chen et al., 7 Oct 2025).

This architecture also imposes a structural claim about conscious unity. Multiple patterns may compete within a given channel at the subconscious level, but at most one pattern per channel appears in a conscious scenario. Consciousness is therefore a sequence of globally integrated channel assignments rather than a mere collection of local feature activations. This is the point at which CER departs from simple feature-map theories: feature maps supply candidate contents, whereas consciousness is the selected global configuration of those contents (Chen et al., 7 Oct 2025).

2. Entropy functionals and the S2C mechanism

The formal core of CER is a family of entropy-like functionals defined over scenario distributions. The basic building block is “refusal entropy,”

$R(a)=a\log \frac{e}{a},$

with $R(0)=0$. On that basis CER defines a Shannon-style entropy for a scenario distribution $x:P^n\to[0,1]$,

$\mathbf{H}(x)=\sum_{r\in P^n} R(x_{\{r\}})-\log e,$

and then generalizes it to a weighted entropy over subsets of scenarios,

$\mathbf{E}_w(x)=\sum_{S\subseteq P^n} w_S\,R(x_S).$

The weights $w_S\ge 0$ encode constraints, instincts, or learned regularities over scenario subsets. Shannon entropy is recovered as the special case in which only singleton subsets are weighted (Chen et al., 7 Oct 2025).

The weighted entropy has a direct representational meaning. A subset $S\subseteq P^n$ can encode a rule, such as an unwanted configuration or a favored correlation between channels. In the illustrative example, “seeing a house leads to feeling big” is represented by assigning weight to the subset of scenarios in which the vision channel is house while the feel channel is not big. Gradient descent then pushes probability mass away from those disallowed scenarios. Entropy reduction in CER is therefore not generic flattening; it is preference-guided elimination of incompatible scenario structure (Chen et al., 7 Oct 2025).

The decisive selection operator is the S2C interface,

$\sum_{p\in P} x(p)=1$0

where $\sum_{p\in P} x(p)=1$1 is an “infinitesimal zero-sum vector with infinitesimal non-zero values and non-zero derivatives.” Its purpose is to break exact symmetries so that the descent dynamics does not remain trapped at unstable symmetric points. CER defines conscious selection as gradient descent on this functional, with normalization steps that preserve the status of $\sum_{p\in P} x(p)=1$2 as a probability distribution and clip probabilities at zero once a scenario is eliminated (Chen et al., 7 Oct 2025).

The key theorem states that gradient descent of a non-zero weighted general entropy always terminates in one of the determined distributions. This gives CER its characteristic interpretation: the conscious scenario is not one candidate among many equally present contents, but the endpoint of a descent process that removes uncertainty until only one scenario remains. In this sense, CER makes entropy reduction constitutive rather than merely correlative of conscious content (Chen et al., 7 Oct 2025).

3. Feedback, internal simulation, and higher-order consciousness

The short-version paper distinguishes a primitive and a higher-order form of CER. In primitive CER, subconsciousness tracks external input by leaky integration,

$\sum_{p\in P} x(p)=1$3

and the conscious scenario at time $\sum_{p\in P} x(p)=1$4 is the output of the S2C map,

$\sum_{p\in P} x(p)=1$5

At this level, the model already supports recognition and immediate action selection, because the subconscious maintains multiple possibilities while S2C chooses one of them as the conscious outcome (Chen et al., 7 Oct 2025).

Higher-order CER adds a consciousness-to-subconscious feedback pathway, C2S. This feedback returns the just-selected conscious scenario to the subconscious landscape so that it can shape subsequent conscious episodes. At the channel level the update is controlled by a decay factor $\sum_{p\in P} x(p)=1$6, and at the full-scenario level the operator $\sum_{p\in P} x(p)=1$7 applies the update channelwise. The result is a recurrent system in which conscious scenarios are no longer terminal outputs but become part of the substrate from which later scenarios are selected (Chen et al., 7 Oct 2025).

This feedback is the mechanism by which CER explains internal simulation. Once a conscious scenario can be fed back into the subconscious distribution, chains of internally generated scenarios become possible. The short-version paper explicitly associates this with dreams, planning, and “thought experiments,” arguing that internal solution of problems is more economical than overt trial-and-error in the world. Higher-order consciousness thus differs from primitive consciousness not by introducing a new substrate, but by recursively reusing selected conscious scenarios as inputs to later selection steps (Chen et al., 7 Oct 2025).

The same framework is extended to subjectivity. “Self” and subjective experiences are treated as internal patterns that can appear in scenarios, and subjective consciousness is the case in which internal C2S loops generate patterns representing the self and its relation to experience. CER therefore places self-consciousness inside the same scenario dynamics as perception and action: self-related content is another class of pattern that can be stabilized by entropy-reducing selection (Chen et al., 7 Oct 2025).

4. Relations to global workspace, IIT, feature binding, and attractor models

CER is presented as a mathematically explicit answer to a problem left under-specified in several adjacent theories. Relative to GWT, subconscious scenario distributions play the role of the many locally generated unconscious contents, while the conscious scenario corresponds to the globally selected content. The S2C mechanism functions as the missing competition-and-selection procedure that decides which content enters the global workspace. In that respect CER is less a rejection of GWT than a proposed implementation of one of its central transitions (Chen et al., 7 Oct 2025).

Relative to IIT, CER agrees that integration matters because correlations across channels reduce entropy. The difference is architectural. IIT measures consciousness by integrated information and treats entropy reduction as a consequence of irreducibility across parts, whereas CER identifies the conscious content with the endpoint of a descent process that drives a scenario distribution to a determined state. The two frameworks therefore share an anti-modular emphasis on cross-channel dependence, but they disagree on whether consciousness is primarily a scalar property of a system or a selection process over scenarios (Chen et al., 7 Oct 2025).

CER also inherits the language of feature binding from Feature Integration Theory. Channels behave like feature maps, and conscious scenarios are the integrated combinations of channelwise patterns. But CER adds a probabilistic scenario space and a descent mechanism. A feature map on its own is only an input to the world of possible subconscious scenarios; consciousness requires the global binding of those features into one selected scenario (Chen et al., 7 Oct 2025).

The short-version paper explicitly connects CER to Hopfield networks. In a Hopfield model, a quadratic energy over binary activations is minimized until the network reaches an attractor. CER preserves the attractor logic but replaces the quadratic energy with a weighted entropy over entire scenarios, and it performs descent in a scenario space whose dimension grows exponentially with the number of channels. This makes the analogy structurally clear while also highlighting CER’s main computational difficulty: the state space is combinatorially large (Chen et al., 7 Oct 2025).

5. Entropy in neighboring frameworks

CER is only one member of a broader research landscape in which “entropy” refers to several different objects. Some neighboring accounts support CER-like readings, but others redefine the relevant entropy or relocate consciousness to a different level of description. A thermodynamic framework of consciousness, for example, explicitly rejects the claim that consciousness is entropy reduction in a simple literal way and instead proposes two non-equilibrium thermodynamic conditions, TCC1 and TCC2, centered on dissipation-driven adaptation, memory $\sum_{p\in P} x(p)=1$8, and low

$\sum_{p\in P} x(p)=1$9

together with low fluctuation term $P^n$0. In that framework, consciousness is better described as efficient homeostatic entropy management than as simple entropy reduction (Ganesh, 2020).

Other nearby theories are more directly CER-compatible. Resonance Complexity Theory (RCT) defines a field entropy

$P^n$1

and states that “CI peaks correspond to phases of maximal integration and reduced entropy,” with low $P^n$2 indicating focused excitation and increased coherence in attractor states (Bruna, 26 May 2025). In a different register, a theory of creative cognition defines “psychological entropy” as arousal-provoking uncertainty and describes creative cognition as recursive restructuring until arousal dissipates and entropy reaches an acceptable level; the paper does not offer a full theory of consciousness, but it provides a cognition-level entropy-minimization template often taken to be CER-adjacent (Gabora, 2016). A later cross-domain synthesis likewise states that “one of the chief functions of consciousness is the transition from a state of uncertainty to a state of certainty,” linking consciousness to surprise resolution and uncertainty reduction across physics, biology, and psychology (Sverdlik, 10 Feb 2026).

The heterogeneity of these usages is clearer when the principal entropy objects are laid side by side.

Framework	Entropy object	Relation to consciousness
CER (Chen et al., 7 Oct 2025)	Scenario entropy over $P^n$3	Conscious content is the zero-entropy, determined scenario selected by S2C
Thermodynamic conditions (Ganesh, 2020)	$P^n$4, heat dissipation, $P^n$5, $P^n$6	Consciousness requires dissipative adaptation and efficient homeostatic entropy management
RCT (Bruna, 26 May 2025)	Field entropy $P^n$7	High-CI attractors correspond to maximal integration and reduced entropy
Transfer-entropy model of DOC (Mäki-Marttunen et al., 2013)	$P^n$8	Consciousness is associated with preserved directed entropy reduction across regions
Info-structural computational model (Iovane et al., 2022)	Shannon entropy over graph weights plus graph energy	Consciousness is indexed by movement in an entropy-energy plane, not by monotonic entropy reduction

This comparison shows that CER’s claim is strongest when entropy is understood as entropy of a scenario distribution or of uncertainty-bearing internal states. It is weaker when entropy is taken to mean thermodynamic entropy production, signal entropy, or graph entropy, because those neighboring literatures often report different monotonicities.

6. Empirical tensions, reinterpretations, and limitations

The largest challenge to CER comes from empirical work in which more conscious states exhibit higher, not lower, entropy-like quantities. A non-equilibrium study of ECoG and fMRI reports that reduced-consciousness states unfold at higher proximity to equilibrium than conscious wakefulness, as shown by entropy production and the curl of probability flux in phase space (Perl et al., 2020). A related multivariate Ornstein–Uhlenbeck analysis of human fMRI finds a monotonic relationship between entropy production and consciousness level during the transition from wakefulness to deep sleep, with entropy production highest in wakefulness and decreasing through N1, N2, and N3 sleep (Gilson et al., 2022). These results place simple thermodynamic CER in direct tension with the non-equilibrium literature.

Signal-entropy and configurational-entropy results generate a similar tension. In EEG, iEEG, and MEG, permutation entropy and permutation Lempel–Ziv complexity tend to be greatest in fully alert states and lower in sleep stages and epileptic seizures (Mateos et al., 2017). At the network level, conscious wakefulness has also been described as the state with the greatest number of possible configurations of interactions between brain networks, representing the highest entropy values (Erra et al., 2016). These findings do not deny that consciousness involves structure, but they do deny that all empirically relevant entropy measures decrease with consciousness.

A further complication comes from the recent “entropy-content conundrum.” The Complex Brain Hypothesis argues that Minimal Phenomenal Experiences and high-content psychedelic experiences may both show elevated brain entropy, even though their phenomenal richness differs sharply. On that basis the paper proposes that richness is better indexed by complexity than by entropy, and that minimal-content and high-content states can both be associated with elevated brain entropy while differing in grain of inference and perturbational signatures (Mago et al., 15 May 2026). In contrast, a DOC study based on transfer entropy remains more congenial to CER, because transfer entropy explicitly measures conditional entropy reduction between regions and is disrupted in patients with disorders of consciousness, while correlating with CRS-R scores and communication function (Mäki-Marttunen et al., 2013).

These tensions suggest that CER is most coherent when interpreted as a theory of representational scenario selection rather than as a universal law about every entropy metric used in consciousness research. The short-version CER paper itself leaves several major questions open: biological realism is not specified; scenario space grows exponentially with the number of channels; the learning of the weight function $P^n$9 is not specified; long-term memory mechanisms require extension; and the hard problem is not solved by the formalism (Chen et al., 7 Oct 2025). As a result, CER currently functions less as a settled empirical doctrine than as a precise proposal about one possible mechanism by which conscious contents could emerge from subconsciously maintained uncertainty.