EvoBrain: Evolving Brain-like Systems

• EvoBrain is a research motif that encompasses the evolution and self-organization of brain-like systems through adaptive, multi-scale processes.
• It spans diverse applications including global brain theories, evolvable robot controllers, neural architecture search for AI, and dynamic EEG modeling for seizure analysis.
• Key themes include modular design, adaptive learning mechanisms, and the interplay of evolutionary and neural processes that enhance performance and interpretability.

EvoBrain denotes a family of research programs centered on the evolution, adaptation, or explicit modeling of brain-like systems across several scales. In the literature summarized here, the term spans at least four distinct usages: a socio-technological “global brain” emerging from ephemeralization and stigmergy; evolvable robot controllers and co-designed bodies and brains; brain-inspired neural substrates and search procedures for artificial intelligence; and EEG-centered models for seizure analysis and continual brain decoding [0703004; (Kotoge et al., 19 Sep 2025, Zhou et al., 1 Jun 2026)]. A plausible implication is that “EvoBrain” functions less as a single canonical architecture than as a recurrent organizing idea: intelligence is treated as something that evolves, self-organizes, or is progressively structured through interaction among modules, environments, and adaptive mechanisms.

1. Conceptual genealogy: from stigmergy to the global brain

One influential conceptual precursor frames technological and social evolution as a trial-and-error process that produces a progressive accumulation of knowledge [0703004]. In that account, technological evolution leads to ephemeralization, defined as ever increasing productivity, or decreasing of the friction that normally dissipates resources. As a result, flows of matter, energy and information circulate ever more easily across the planet. Increased global connectivity raises interactions between agents and therefore possibilities for conflict, but evolutionary progress is also described as reducing social friction via the creation of institutions, conceived as “mediators” [0703004].

The same framework assigns a central role to stigmergy, the unintended collaboration between agents resulting from their actions on a shared environment. The Internet is described as a near ideal medium for stigmergic interaction. Two forms are distinguished. Quantitative stigmergy allows the web to learn from the activities of its users, thus becoming ever better at helping them to answer their queries. Qualitative stigmergy stimulates agents to collectively develop novel knowledge. Both are said to have direct analogues in the functioning of the human brain, motivating the vision of a future, super-intelligent web as a “global brain” for humanity [0703004].

This literature also links social and technological feedback to an extreme acceleration of innovation, and an extrapolation of the corresponding hyperbolic growth model is described as forecasting a singularity around 2040, interpreted as the evolutionary transition to the Global Brain regime [0703004]. This suggests an early, high-level sense of EvoBrain in which “brain” refers not to a bounded artifact but to planetary-scale collective cognition emerging from reduced friction, mediated coordination, and shared informational environments.

2. EvoBrain in embodied evolution and robot control

In evolutionary robotics, EvoBrain refers more concretely to the evolution or co-optimization of robot controllers together with morphology. One line of work compares brain reproduction mechanisms in systems where both bodies and brains evolve and the brains of “infant” robots are optimized by a learning process immediately after “birth” (Luo et al., 2023). Four combinations are studied by crossing asexual or sexual brain reproduction with Darwinian or Lamarckian inheritance. The reported result is that sexual reproduction of the robots’ brains is preferable in the Darwinian framework, but the effect is the opposite in the Lamarckian system; the overall best option is reported as asexual reproduction combined with the Lamarckian framework, while the evolved morphologies do not differ across brain reproduction methods (Luo et al., 2023). The interpretation offered in the source is mechanistic: Lamarckian systems write learned, body-specific parameter structures back into the genotype, and sexual crossover can then disrupt coherent morphology-aligned weight fields.

A more recent co-design formulation is "ECo-MoE: Embodiment-Conditioned Mixture of Experts Increases the Evolvability of Robots" (Wang et al., 22 May 2026). Here the “genotype” is a latent vector $z \in \mathbb{R}^{512}$ sampled from an evolving Gaussian distribution, and the “phenotype” is a decoded body plan $x = D(z)$ represented as a $64 \times 64 \times 64$ voxel tensor with soft tissue and rigid bone channels and bone IDs. Control is produced by a mixture of expert policies whose weights are gated by the latent design coordinates,

$$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$

The method is explicitly positioned between two extremes: training an individual policy for every robot, which is inefficient, and a monolithic universal controller, which results in overly conservative structures and behaviors (Wang et al., 22 May 2026).

The optimization scheme separates timescales. Evolution uses CMA-ES to update the latent design distribution, while PPO updates the experts and the gate. A routing diversity regularizer is added to prevent expert collapse. Empirically, the method is reported to significantly increase evolvability versus a single-expert baseline on Upright Locomotion and Potholes, while Flat Ground shows similar performance because both approaches converge to fast “snake” morphologies. Parameter efficiency is also reported: baseline total policy size $2.563\text{M}$ versus ECo-MoE total $2.388\text{M}$, with $0.597\text{M}$ per expert for four experts and a shared critic identical to baseline at $2.038\text{M}$ (Wang et al., 22 May 2026).

Another strand emphasizes intelligibility rather than modular expert reuse. "Behaviour Trees for Evolutionary Robotics" applies the Behaviour Tree framework to a real robotic platform using the Evolutionary Robotics methodology (Scheper et al., 2014). The key claim is that Behaviour Trees improve the intelligibility of emergent robotic behaviour relative to traditional neural network formulations and thus make manual adaptation across the reality gap more tractable. On the DelFly Explorer flapping wing MAV, the evolved Behaviour Tree achieved $88\%$ success in simulation and $54\%$ in real-world flight tests after user adaptation, compared with $x = D(z)$0 for a tuned user-defined controller (Scheper et al., 2014). In this usage, an EvoBrain is not only evolvable but also hierarchically inspectable.

3. Evolvable neural substrates and brain-inspired architectures

A more literal meaning of EvoBrain appears in work that evolves computational substrates resembling neural organization. "Brain-inspired Evolutionary Architectures for Spiking Neural Networks" proposes Evolutionary Brain-inspired Neural Architecture Search (EB-NAS) for SNNs (Pan et al., 2023). The search space has local modules composed from five canonical motifs—Feedforward Excitation, Feedforward Inhibition, Feedback Inhibition, Lateral Inhibition, and Mutual Inhibition—and global directed feedforward and feedback connectivity among modules, with self-loops disallowed. Architectures are optimized by NSGA-II with a few-shot CART-based surrogate predictor, using a two-objective formulation

$x = D(z)$1

where $x = D(z)$2 is validation error or loss and $x = D(z)$3 counts spikes as an energy proxy (Pan et al., 2023).

The reported results are dataset-specific and concrete: $x = D(z)$4 top-1 accuracy on CIFAR10 at $x = D(z)$5, $x = D(z)$6 on CIFAR100 at $x = D(z)$7, $x = D(z)$8 on CIFAR10-DVS at $x = D(z)$9, and $64 \times 64 \times 64$0 on DVS128-Gesture at $64 \times 64 \times 64$1 (Pan et al., 2023). On CIFAR10, the method is also reported to achieve $64 \times 64 \times 64$2 at $64 \times 64 \times 64$3 spikes, compared with AutoSNN at $64 \times 64 \times 64$4 and $64 \times 64 \times 64$5 spikes. This line makes EvoBrain synonymous with evolutionary discovery of modularity, E/I structure, and long-range connectivity under joint accuracy–energy constraints.

At a finer granularity, "Network of Evolvable Neural Units: Evolving to Learn at a Synaptic Level" evolves individual soma and synaptic compartment models rather than network topology alone (Bertens et al., 2019). Each neuron and each synapse is an Evolvable Neural Unit with shared gate matrices and local dynamic state:

$64 \times 64 \times 64$6

Using Evolution Strategies with $64 \times 64 \times 64$7 offspring per generation, $64 \times 64 \times 64$8, $64 \times 64 \times 64$9, and momentum $$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$0, the framework is reported to evolve units that mimic Integrate-And-Fire neurons, reproduce a classic STDP timing curve, and support one-shot learning in a T-maze through locally evolved reward-based synaptic dynamics (Bertens et al., 2019). In this sense, EvoBrain refers to bottom-up evolution of neural computation and plasticity rules.

"Markov Brains: A Technical Introduction" provides another canonical evolvable-brain substrate (Hintze et al., 2017). Markov Brains are networks of computational components, or gates, operating on a state buffer of sensor, hidden, and motor nodes. Gates synchronously read from the “old” buffer and write to a separate “new” buffer, with multiple writes to the same node combined by summation. The genome is circular and parsed left-to-right; specific start codons mark genes, including $$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$1 for deterministic logic gates and $$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$2 for probabilistic logic gates. In probabilistic gates, a matrix $$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$3 specifies $$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$4 with row normalization (Hintze et al., 2017). This work makes explicit that an evolvable brain can be heterogeneous, gate-based, and only loosely analogous to conventional layered neural networks.

A broader cognitive-architectural formulation appears in "Design for a Darwinian Brain: Part 2. Cognitive Architecture" (Fernando et al., 2013). There, actors implement behaviors and games implement task or fitness functions, both co-evolving while interacting through working memory and long-term memory. Actor fitness aggregates across associated games,

$$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$5

whereas game fitness is the variance of actor performance on that game,

$$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$6

This architecture treats open-ended learning as co-evolution of behaviors and self-generated objectives, embedding yet another meaning of EvoBrain: an evolvable physical symbol system coupled to an evolvable ecology of tasks (Fernando et al., 2013).

4. EEG-specific EvoBrain: dynamic graph modeling and continual decoding

The term EvoBrain is also used directly for EEG models. "EvoBrain: Dynamic Multi-channel EEG Graph Modeling for Time-evolving Brain Network" addresses seizure detection and early seizure prediction by explicitly modeling time-varying connectivity among EEG channels (Kotoge et al., 19 Sep 2025). The work states two theoretical results. First, explicit dynamic graph modeling, in which adjacency matrices vary with time, is strictly more expressive than implicit dynamic modeling with a fixed adjacency:

$$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$7

Second, among three node-level dynamic GNN patterns under $$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$8-WL graph encoders, graph-then-time is strictly less expressive than time-and-graph, which is strictly less expressive than time-then-graph:

$$w(z) = \operatorname{softmax}(G(z)),\quad \pi(a\mid s,z) = \sum_{k=1}^K w_k(z)\,\pi_k(a\mid s).$$9

The proposed model follows the most expressive pattern, combining a two-stream Mamba with a GCN enhanced by Laplacian positional encoding (Kotoge et al., 19 Sep 2025).

The architecture builds a dynamic adjacency at each snapshot using normalized cross-correlation on frequency-domain node features, retaining the top-$2.563\text{M}$0 neighbors per node. Temporal evolution of nodes and edges is encoded separately with a selective state-space model,

$2.563\text{M}$1

$2.563\text{M}$2

Spatial reasoning is then performed once, after temporal aggregation, by a two-layer GCN with $2.563\text{M}$3 hidden units. On TUSZ, the reported performance is AUROC $2.563\text{M}$4 and F1 $2.563\text{M}$5 for $2.563\text{M}$6 seizure detection, and AUROC $2.563\text{M}$7 and F1 $2.563\text{M}$8 for $2.563\text{M}$9 seizure prediction. The paper also reports that EvoBrain is up to $2.388\text{M}$0 faster in training time and $2.388\text{M}$1 faster in inference time than DCRNN, because graph convolution is applied only once after temporal aggregation (Kotoge et al., 19 Sep 2025).

A second direct usage is "EvoBrain: Continual Learning of EEG Foundation Models Across Heterogeneous BCI Tasks" (Zhou et al., 1 Jun 2026). Here EvoBrain denotes a continual-learning framework for adapting a single pretrained EEG foundation backbone across a sequence of heterogeneous tasks. Two components are central. Neuro-Spectral Task Normalization (NSN) combines Prototype-guided Temporal Normalization with Spectral Residual Modulation. Its temporal normalization is written as

$2.388\text{M}$2

with target moments interpolated between current-task and historical prototype statistics. Spectral modulation uses six EEG bands—$2.388\text{M}$3—and learnable band gates $2.388\text{M}$4 (Zhou et al., 1 Jun 2026).

The second component, Response-Affinity Distillation (RAD), combines Response-Geometry Distillation with Spectral-Affinity Alignment:

$2.388\text{M}$5

Replay is weighted toward older tasks using

$2.388\text{M}$6

Across six downstream tasks and four EEG foundation backbones, the paper reports stable final Average B-ACC across three task orders, including $2.388\text{M}$7 for CBraMod, $2.388\text{M}$8 for EEGMamba, $2.388\text{M}$9 for CSBrain, and $0.597\text{M}$0 for DeeperBrain, with Relative Forgetting Rate ranging from $0.597\text{M}$1 to $0.597\text{M}$2 (Zhou et al., 1 Jun 2026). In this formulation, EvoBrain is a dynamic post-training regime rather than a new EEG encoder.

An earlier EEG optimization pipeline also falls within the same family of meanings. "A Deep Evolutionary Approach to Bioinspired Classifier Optimisation for Brain-Machine Interaction" uses evolution to select from an initial set of $0.597\text{M}$3 EEG statistical features and to optimize MLP topology before training (Bird et al., 2019). The selected subsets are $0.597\text{M}$4 features for mental state, $0.597\text{M}$5 for emotional state, and $0.597\text{M}$6 for MindBigData digits. Reported accuracies include $0.597\text{M}$7 for AdaBoosted LSTM on attention, $0.597\text{M}$8 on emotion, and $0.597\text{M}$9 for AdaBoosted DEvo MLP on digits, with the nonboosted DEvo MLP reaching $2.038\text{M}$0, $2.038\text{M}$1, and $2.038\text{M}$2 respectively (Bird et al., 2019). This makes EvoBrain, in a weaker sense, a bioinspired evolutionary optimization stack for EEG classification.

5. Hardware and biological interfacing

A hardware-oriented meaning appears in "A biologically interfaced evolvable organic pattern classifier" (Gerasimov et al., 2022). The system is a single-layer, hardware perceptron built from evolvable organic electrochemical transistors (EOECTs), where each signed synaptic weight is represented by two EOECTs. With binary inputs, the single-pixel current is

$2.038\text{M}$3

so the effective weight is $2.038\text{M}$4. The classifier is trained in situ by the Widrow–Hoff learning rule,

$2.038\text{M}$5

mapped physically onto conductance increments via electropolymerization (Gerasimov et al., 2022).

The reported hardware properties are specific. Channel conductance ranges from approximately $2.038\text{M}$6 after $2.038\text{M}$7 pulses to approximately $2.038\text{M}$8 after $2.038\text{M}$9 pulses. Across $88\%$0 channels grown with $88\%$1 pulses, read currents at $88\%$2 show $88\%$3 relative standard deviation. Dry-state resistance drift is approximately $88\%$4 over an average of $88\%$5 hours, described as an improvement of two orders of magnitude over ion-trapping neuromorphic OECTs (Gerasimov et al., 2022).

The system is coupled to an organic electrochemical integrate-and-fire neuron, which converts classifier output current into spikes. Trained on a $88\%$6 array to classify the letter “T” versus “J,” the full chain is sensor input $88\%$7 EOECT classification $88\%$8 transimpedance amplifier $88\%$9 organic electrochemical neuron $54\%$0 cuff electrode stimulation of the leech ventral nerve cord. Reported OECN spikes have approximately $54\%$1 amplitude, approximately $54\%$2 frequency, and approximately $54\%$3 half-width for “T” inputs, with no spiking for “J.” Significant contractions are observed only during “T” inputs (Gerasimov et al., 2022). In this literature, EvoBrain is instantiated as physically evolvable synapses and neuromorphic signal translation in direct interface with living tissue.

6. Unifying themes, distinctions, and recurrent debates

Across these literatures, EvoBrain does not designate a single technical object. The strongest commonality is structural rather than architectural: adaptive competence is produced by evolutionary search, local update rules, or explicit modeling of dynamic relations, and performance depends on preserving useful structure while allowing selective change. In robot co-design this appears as embodiment-conditioned routing that localizes specialization (Wang et al., 22 May 2026); in continual EEG learning it appears as the plasticity–stability trade-off handled by NSN, RAD, and replay (Zhou et al., 1 Jun 2026); in dynamic seizure modeling it appears as the explicit evolution of nodes and edges over time rather than reliance on static graphs (Kotoge et al., 19 Sep 2025).

A second recurrent theme is modularity. EB-NAS evolves region-like modules built from canonical motifs and long-range cross-module connections (Pan et al., 2023). Markov Brains are assembled from heterogeneous gates with explicit read–write semantics (Hintze et al., 2017). Darwinian Brain architectures define modular atoms and molecules, with actors and games co-evolving through working memory and long-term memory (Fernando et al., 2013). Behaviour Trees in evolutionary robotics make sub-behaviours inspectable and therefore manually adaptable across simulation and reality (Scheper et al., 2014). This suggests that EvoBrain research repeatedly treats modular decomposition as a precondition for evolvability, interpretability, or both.

A third theme is the status of learning within evolution. Lamarckian inheritance can outperform Darwinian inheritance when learned controller improvements can be written back coherently into a direct encoding, but sexual recombination can then become destructive (Luo et al., 2023). Evolvable Neural Units push this further by evolving the update rules themselves, allowing in-lifetime learning to emerge from synaptic and neuronal state dynamics rather than from an external optimizer (Bertens et al., 2019). Hardware EOECT classifiers similarly treat synaptic change as a physical process enacted in situ (Gerasimov et al., 2022). By contrast, some EEG uses of EvoBrain refer primarily to architectural or post-training organization rather than to open-ended evolutionary search (Kotoge et al., 19 Sep 2025, Zhou et al., 1 Jun 2026).

The main misconception is therefore that EvoBrain names one established paradigm with a fixed model class. The record summarized here does not support that reading. Instead, the term encompasses a spectrum from conceptual global-brain theory to evolvable robot controllers, neural substrates, EEG graph models, continual-learning frameworks, and organic biointerfaces [0703004; (Kotoge et al., 19 Sep 2025, Zhou et al., 1 Jun 2026)]. A plausible implication is that the encyclopedia meaning of EvoBrain is best given as a research motif: the progressive construction of brain-like intelligence through evolution, adaptive interaction, or explicit temporal–structural organization, rather than a single standardized method.