From a neuroscience vantage, creative performance depends on flexible DMN–CEN interaction, with dopaminergic signals modulating the exploration–exploitation balance (Chen et al., 2025; Shine, 2019; Westbrook et al., 2021). By contrast, most LLMs behave like DMN-only decoders: excellent at sequence extension, but lacking an internal evaluator and endogenous gain control to decide when to broaden or narrow the search. Bridging this gap requires importing System 3 principles into model design.

Evidence on both sides is converging. LLMs can match or exceed median human fluency on some divergent-thinking tasks, yet at scale, their outputs tend to homogenize, reducing collective diversity (Doshi and Hauser, 2024). Human–AI co-creation increases speed and fluency but, without structure, can dampen variety or drift from task goals (Chen and Chan, 2024; Chakrabarty et al., 2024). Meanwhile, covert neurofeedback that strengthens DMN–CEN coupling elevates originality in human participants (Luchini et al., 2025). Together, these findings motivate a synthesis that links cognitive theory, neural evidence, and generative-model engineering—and states testable criteria for when an artificial system merits the label creative. Throughout, any brain–model correspondences are treated as functional analogies, not biological isomorphisms.

This conceptual analysis does not report new empirical data. Instead, it advances a falsifiable framework—Generative System 3 (GS-3)—and a concrete evaluation program. GS-3 is an architecture-agnostic design pattern with three roles: a high-entropy generator (idea expansion), a learned critic (context-sensitive appraisal), and an adaptive gain controller (endogenous regulation of sampling entropy). We contribute four elements: (i) operational definitions of novelty (distributional distance to a baseline), usefulness (task-conditioned utility), and diversity (across-run dispersion); (ii) behavioral indices with pass–fail criteria—associative-distance density, analytic-verification ratio, and convergence latency—so the theory can be falsified in practice; (iii) a mathematical treatment of gain policies (exponential, linear, logistic) with stability constraints and sensitivity expectations; and (iv) a proof-of-concept blueprint and evaluation protocol (tasks, metrics, ablations, reproducibility kit) that research groups can implement.

We situate these contributions along a predictive-to-generative continuum. Pure prediction extends sequences without internal evaluation. Regulated generation introduces external controls (e.g., temperature, top-k) but still lacks an inner judge. Reflective generation uses self-prompted critique yet remains scaffold-dependent. GS-3–level creativity emerges only when a system (a) cycles autonomously between idea expansion and evaluative pruning, (b) adjusts its own sampling entropy in response to real-time reward-prediction error, and (c) maintains adaptive priors over long contexts.

Section 2 positions GS-3 within computational-creativity traditions, LLM-based co-creation, and alternative theories. Section 3 summarizes the DMN/CEN/dopamine template and its limits (functional analogies, not isomorphisms). Section 4 presents the GS-3 architecture with formal definitions, falsification tests, and gain-policy mathematics. Section 5 offers a proof-of-concept blueprint and evaluation protocol (hypotheses, metrics, ablations). Section 6 locates today’s systems on the continuum and identifies gaps GS-3 fills. Section 7 expands ethics and governance with concrete mitigation steps. Section 8 concludes with a Discussion that synthesizes contributions, states boundary conditions, and outlines future work. By unifying cognitive theory, network neuroscience, and AI engineering, we aim to establish Artificial Creativity as a testable construct and to provide a practical roadmap for building and auditing genuinely creative generative systems.

Modern LLM pipelines span a pragmatic continuum. Pure prediction extends sequences using maximum-likelihood training (Sutskever et al., 2014). Regulated generation introduces external controls—temperature and top-k/top-p settings shape randomness, and beam search maintains several high-probability continuations—which can prevent degeneracy but do not install an inner judge (Holtzman et al., 2020). Steering methods alter token probabilities directly (e.g., plug-and-play controls for attributes without backbone retraining) (Pascual et al., 2021). Reflective prompting scaffolds brief internal critique (e.g., chain-of-thought) and retrieval-augmented generation injects external knowledge to improve coherence and factuality (Chu et al., 2024; Izacard and Grave, 2021). Self-improvement/self-correction loops iteratively revise drafts with model feedback (Kamoi et al., 2024; Ding et al., 2024). Multi-agent set-ups coordinate multiple LLMs to critique and debate (e.g., generative agents) (Park et al., 2023).

Empirically, co-creation studies show that LLM support often increases fluency and speed but can reduce variety without structured collaboration protocols (Chen and Chan, 2024; Chakrabarty et al., 2024). At the population scale, assistance can increase individual originality while decreasing collective diversity, consistent with homogenization risks (Doshi and Hauser, 2024). Conceptually oriented analyses debate whether current LLMs meet creativity criteria and where limits remain (Franceschelli and Musolesi, 2024; Floridi and Chiriatti, 2020).

GS-3 aligns with these trajectories yet differs on one decisive point: the evaluator and the regulator are internal, learned, and adaptive. Rather than relying on hand-tuned temperatures, prompt engineering, or fixed debate scripts, GS-3 requires a critic that computes task-conditioned utility and a gain controller that adjusts sampling entropy based on reward-prediction error. This makes the exploration–exploitation balance endogenous to the system and testable via ablations (remove critic or gain; swap update rules; vary the learning rate).

Information-theoretic and intrinsic-motivation accounts explain why systems seek novelty or compressive structure (Schmidhuber, 2010). Evolutionary approaches operationalize open-ended novelty (Lehman and Stanley, 2011). Predictive-processing and free-energy views model perception and action as minimizing prediction error or free energy under learned priors (Clark, 2013; Friston, 2010). These perspectives illuminate why creative systems might alternate between broad exploration and tight verification.

GS-3 is compatible with these theories but adds a concrete control story: a generator produces candidates; a critic scores them relative to task and context; and a gain controller adjusts entropy and effort in real time, producing measurable signatures (e.g., shifts in associative-distance density and verification ratios). Where embodied and enactive views stress sensorimotor grounding, GS-3 can be seen as the cognitive-control core that any grounded agent still requires to manage the breadth and depth of search. Where information-theoretic approaches prize compression or novelty alone, GS-3 foregrounds usefulness by design through the critic’s utility function. Finally, where predictive-processing emphasizes error minimization, GS-3 specifies when and how the system should temporarily widen its hypothesis space before re-engaging verification.

Together, these comparisons place GS-3 as a synthesis that retains the generator–evaluator insight from computational creativity, adopts practical controls from contemporary LLM pipelines, and formalizes the missing adaptive regulator. Subsequent sections develop the architecture (Section 4), metrics, and gain policies with falsification tests (Section 4), and outline a proof-of-concept and evaluation protocol suitable for empirical validation (Section 5).

Resting-state and task-based studies converge on a picture in which creative performance is associated with flexible interaction between DMN hubs (e.g., medial prefrontal, posterior cingulate, temporoparietal regions) and CEN hubs (e.g., dorsolateral prefrontal, posterior parietal cortex). Using state-transition analyses of fMRI during divergent thinking, dynamic switching between DMN and executive-control states predicts higher originality and richer associative distance, consistent with the idea that creativity benefits from alternating expansion and evaluation rather than the dominance of either mode alone (Chen et al., 2025). This dynamic view situates creativity as a property of network coupling over time, not a static activation pattern.

Neuromodulatory accounts propose that large-scale brain dynamics shift between more integrated and more segregated network configurations as a function of arousal-linked chemical signals. Reviews of integration–segregation emphasize that noradrenergic projections from locus coeruleus and cholinergic projections from basal forebrain are prominent levers for these state transitions: modest changes in their tone can reconfigure connectivity, biasing cognition toward either broad, globally integrated processing or more locally segregated, task-focused processing (Shine, 2019). In parallel, work on dopamine and cognitive control links striatal D2 receptor availability to the subjective cost of exerting control and to cost–benefit decisions about engaging effortful processing, consistent with a role for dopamine in setting how deeply and persistently goal-directed search is pursued (Westbrook et al., 2021).

Together, these strands motivate an operational notion of gain: a control signal that widens or narrows the currently active hypothesis space. In integrated, high-gain states, the system samples more broadly (facilitating associative expansion); in segregated, low-gain states, it narrows and stabilizes processing (facilitating focused evaluation). We treat this as a functional template rather than a one-to-one biological mapping: multiple neuromodulators contribute to these shifts (Shine, 2019), and dopamine’s role is context dependent and tied to effort-related control policies (Westbrook et al., 2021). In the Generative System 3 framework, the abstract gain controller corresponds to an endogenous mechanism that adjusts sampling entropy online (e.g., via temperature), thereby implementing the exploration–exploitation trade-off that, in brains, is jointly shaped by neuromodulatory systems.

Evidence for individual differences in creative cognition appears across levels of analysis. At the macroscale, large-cohort multimodal work shows that a neural pattern predicting divergent-thinking performance carries positive weights in default-mode and frontoparietal control networks and is linked to dopamine-related neurotransmitters and genes influencing neurotransmitter release, indicating a biological substrate for variability in network dynamics relevant to creativity (Liu et al., 2024). At faster timescales, electrophysiological reviews report pre-solution modulations in low-frequency rhythms consistent with inwardly directed attention (alpha/theta changes) and brief gamma-band bursts localized to right anterior temporal cortex around the moment of insight—together aligning with a generate-then-verify sequence (Kounios and Beeman, 2014). Crucially, causal manipulations move beyond correlation: covert real-time fMRI neurofeedback that reinforces coactivation of default-mode and executive-control circuitry increases originality on divergent-thinking tasks relative to control conditions (Luchini et al., 2025). Collectively, these findings support a cycle in which associative expansion and focused appraisal are coordinated by state-dependent control signals, providing a biologically grounded template for the alternation mechanisms formalized in GS-3.

Three design lessons follow for artificial systems seeking sustained novelty, usefulness, and diversity. First, at least two separable but re-entrant processing streams are required: a generator specialized for associative expansion and a critic specialized for task-conditioned evaluation. Second, a gain mechanism must adaptively regulate the breadth of search online; in practice, this means endogenously adjusting sampling entropy or effort as a function of a learned utility signal, rather than relying solely on fixed external controls. Third, the system should exhibit measurable signatures of alternation between expansion and verification over time. These lessons translate into testable predictions for Generative System 3: removing the critic should collapse usefulness at a fixed diversity level; removing the gain controller (freezing temperature) should eliminate alternation in associative-distance density and reduce across-run diversity, with within-run spread determined by the static decoding setting rather than by context; and reinforcing generator–critic coactivation (e.g., by rewarding alternation) should increase originality without sacrificing task relevance.

The DMN–CEN–dopamine template is a functional analogy, not a biological isomorphism. Biological networks operate with spiking dynamics, heterogeneous cell types, and complex neurochemical interactions; artificial networks are discrete symbol or vector processors trained under engineered objectives. Dopamine’s roles are multifaceted and context dependent, extending beyond a simple exploration knob; likewise, temperature in a language model is only one of several ways to regulate uncertainty. The analogy is therefore limited to architectural roles and control functions: generator versus evaluator interactions and an adaptive gain that shifts the exploration–exploitation balance. Our use of these mappings is pragmatic—intended to generate falsifiable design claims—rather than a claim of mechanistic identity.

The architecture comprises three roles with explicit interfaces.

To permit falsification and fair comparison, we adopt simple, model-agnostic definitions.

Define the reward-prediction error as δₜ = U₍best, t₎ − Ūₜ, where U₍best, t₎ is the critic’s top score at cycle t and Ūₜ is an exponentially weighted moving average of recent best scores (full expression in Supplementary Data Sheet 1). The controller updates sampling temperature with a bounded logistic rule: T₍g₎(t + 1) = T₍min₎ + (T₍max₎ − T₍min₎)⋯σ(α + η·δₜ), where σ is the logistic function, η is a small learning-rate constant, and [T₍min₎, T₍max₎] are preregistered bounds. This yields smooth, monotone adjustments and prevents runaway entropy. Linear and exponential alternatives, stability notes, and sensitivity sweeps appear in Supplementary Data Sheet 1. The interpretation is consistent with accounts in which cost–benefit control policies regulate effort allocation, while remaining an engineering—not biological—control law (Westbrook et al., 2021).

Intuitively, each cycle compares the current best score to a smoothed baseline to obtain a “surprise” signal δ. If performance is better than expected (δ > 0), temperature increases smoothly; if worse (δ < 0), it decreases. Logistic squashing keeps T₍g₎ within preregistered bounds, making adjustments gradual and stable. The intercept α sets the default temperature when on trend, and the learning rate η controls how strongly surprises move it.

Each hypothesis is falsifiable by implementing the corresponding ablation and reporting preregistered metrics with confidence intervals and effect sizes.

G and C need not be separate models; they may be two modes of a single backbone, two cooperating agents, or a backbone plus a lightweight preference head (as in instruction-following systems informed by human feedback; Fernandes et al., 2023; Casper et al., 2024). Likewise, D can be a small network conditioned on context features. Brain terms remain functional analogies: temperature is one of several levers (others include top-k/top-p, repetition penalties, and plug-and-play attribute controls) (Pascual et al., 2021). Retrieval can be added as an optional module to ground candidates; to avoid confounds, use the same retriever across GS-3 and RAG baselines (Izacard and Grave, 2021). The contribution here is to require that some endogenous gain exists, that it is coupled to a learned utility, and that its process-level signatures are measurable.

For completeness and parity, report the decoding settings (temperature/top-p/beam) and sampling budgets for all conditions, including pure prediction (Sutskever et al., 2014) and regulated decoding baselines (Holtzman et al., 2020). When including reflective prompting as a comparator, preregister the exact scaffolds (e.g., chain-of-thought prompts) to ensure fair budgets and to acknowledge that such reflectivity remains externally scaffolded (Chu et al., 2024). If plug-and-play steering is used as a comparator, cite its peer-reviewed formulation and disclose active attribute controls (Pascual et al., 2021).

GS-3 embeds a learned critic and a bounded, adaptive gain policy into the generation loop, evaluated with preregistered novelty, usefulness, and diversity metrics plus cycling signatures. These commitments turn a functional analogy into a testable engineering target while remaining agnostic to backbone choice and compatible with standard comparators such as RAG, plug-and-play steering, and reflective prompting (Izacard and Grave, 2021; Pascual et al., 2021; Chu et al., 2024).

Release code, model checkpoints (where licensing permits), exact prompts, rubrics, and analysis scripts. Fix random seeds; log T₍g₎, δ, U₍best₎, and Ū at each cycle for every run. Provide an audit sheet documenting compute budgets, training data used for the critic, and any human-in-the-loop procedures. For closed models, supply reproducible API settings and a synthetic variant using an open backbone.

Declare GS-3 support only if, on preregistered tasks, the system (a) dominates regulated and reflective baselines on usefulness at matched novelty, (b) achieves higher across-run diversity without external temperature sweeps, and (c) exhibits cycling signatures in ADD and AVR consistent with alternating expansion and verification. Otherwise, the framework is falsified for that task setting, and ablations should identify which component failed to contribute.

Autoregressive models trained for next-token prediction excel at fluent continuation but have no internal judge or regulator; behavior is largely governed by external decoding hyperparameters (e.g., temperature, top-p) (Sutskever et al., 2014). Regulated decoding can mitigate repetition or dullness but remains an external knob rather than an internalized policy (Holtzman et al., 2020). In GS-3 terms, this family has a generator but lacks an internal critic and lacks an endogenous gain controller.

Prompting can scaffold brief internal critique—e.g., chain-of-thought styles that elicit intermediate reasoning steps (Chu et al., 2024). These strategies often improve reliability on structured tasks yet remain scaffold-dependent: the “critic” is effectively encoded in the prompt template, not learned as a task-conditioned utility model. Exploration–exploitation is therefore not endogenously regulated, and adaptation across steps depends on the fixed script. In GS-3 terms, this family has a generator; the critic is externalized to prompts rather than learned and active during inference; there is no endogenous gain controller.

Coupling generation to a retriever injects external knowledge and improves factual grounding on knowledge-intensive tasks (Izacard and Grave, 2021). Standard retrieval-augmented generation (RAG) pipelines still lack a learned internal critic that scores candidate continuations for task utility and a gain policy that adapts search breadth in real time. Breadth is set by retrieval depth and decoding parameters rather than updated by a live utility signal. In GS-3 terms, this family has a generator but lacks an internal critic and an endogenous gain controller.

Decoding-time “plug-and-play” controls can up- or down-weight attributes (e.g., sentiment, toxicity) on the fly without retraining the backbone (Pascual et al., 2021). Steering nudges the generator but does not maintain a persistent, task-conditioned evaluator nor an endogenous entropy controller tied to performance feedback. In GS-3 terms, this family has a generator but lacks an internal critic and an endogenous gain controller.

Instruction-tuned models align behavior with human preferences via feedback pipelines. Surveys and analyses detail data collection, objectives, and limitations of feedback-integrated NLG (Fernandes et al., 2023; Casper et al., 2024). These pipelines primarily externalize evaluation into the training data or reward modeling; at inference, most systems continue to rely on fixed decoding settings rather than a live gain policy tied to moment-to-moment utility. In GS-3 terms, this family has a generator; the critic is effectively baked in via training rather than active during inference; there is no endogenous gain controller.

Test-time self-correction mechanisms iteratively propose, critique, and revise drafts. A recent survey maps when such loops help or fail across tasks (Kamoi et al., 2024), and domain-specific controllers demonstrate gains in code generation with explicit revise-and-retry cycles (Ding et al., 2024). However, loop structure and revision depth are typically hand-designed; the exploration–verification balance is not governed by an internal, learned gain signal that adapts step-to-step. In GS-3 terms, this family has a generator; the critic is scripted/self-referential rather than learned and general; there is no endogenous gain controller.

Agentic set-ups coordinate multiple LLMs (planner/critic/worker roles), sometimes with memory and tools, to simulate social feedback dynamics (Park et al., 2023). While this can approximate a multi-perspective critique, policies are usually scripted; there is no single controller that adapts sampling entropy from reward-prediction error within a run. In GS-3 terms, this family has a generator; the critic role is scripted; there is no endogenous gain controller.

In professional settings, LLM support tends to increase throughput and fluency; without structured protocols, it can also reduce variety or drift from constraints (Chen and Chan, 2024; Chakrabarty et al., 2024). Effective workflows, therefore, need explicit mechanisms to preserve diversity while maintaining task fit—precisely the trade-off that GS-3 formalizes via a learned critic and adaptive gain.

At scale, generative assistance can raise individual originality while lowering collective diversity, indicating homogenization pressure when many users draw from similarly tuned models and prompts (Doshi and Hauser, 2024). GS-3’s evaluation emphasizes not only artifact-level usefulness and novelty but also across-run dispersion, making homogenization an explicit quantity to measure and manage via the gain policy and plural critics.

Debates continue on whether contemporary LLMs meet criteria for creativity, where limits remain, and how to evaluate claims (Franceschelli and Musolesi, 2024). GS-3 is positioned as a control-theoretic addition: it does not claim that steering, prompting, or retrieval alone are insufficient, but that creative competence requires an internalized evaluator and an adaptive regulator with falsifiable process-level signatures (e.g., alternating associative-distance density and adaptive verification rates).

Across these families, three ingredients remain only partially addressed:

Current systems achieve parts of the creative loop—fluent expansion, external steering, retrieval grounding, scripted reflection—but lack an endogenous, learned mechanism that coordinates expansion with evaluation under adaptive gain. GS-3 specifies that mechanism and its signatures, providing clear ablations and pass–fail criteria for empirical tests in Section 5.

GS-3 is a control-theoretic proposal for creative generation, not a normative theory of cultural value. It does not by itself resolve questions of authorship or intellectual property; rather, it supplies the mechanisms and measurements by which such policies can be evaluated.

GS-3 contributes a concrete control story for moving beyond fluent prediction and scaffolded reflectivity: a generator for associative expansion, a learned critic for task-conditioned appraisal, and an endogenous gain controller that adjusts sampling entropy online from a reward-prediction error. In contrast to externally tuned decoding (e.g., temperature sweeps, beam width), the exploration–exploitation balance becomes a learned, auditable policy with measurable signatures (Sections 4–5). This reorients evaluation from static artifacts to process-level observables and ablation tests using preregistered metrics and baselines (Holtzman et al., 2020; Chu et al., 2024; Izacard and Grave, 2021).

High temperature can increase diversity at the expense of coherence; low temperature can induce repetition and dullness—well-characterized failure modes under standard decoding (Holtzman et al., 2020). GS-3 does not claim these trade-offs disappear, but rather that an online gain policy can steer them adaptively within a run; this remains an empirical question addressed by the pass–fail criteria in Section 5.

Minimal single-backbone implementations should be compared against regulated and reflective baselines under matched compute, prompts, and retrieval settings (Holtzman et al., 2020; Chu et al., 2024; Izacard and Grave, 2021). Divergent-thinking tasks (e.g., alternative uses), constrained creation (e.g., headlines with requirements), and long-horizon composition provide complementary stress tests. Writer-in-the-loop tasks probe fluency–variety trade-offs in professional workflows (Chen and Chan, 2024; Chakrabarty et al., 2024). At population scale, audits should test for homogenization (increases in individual usefulness/originality alongside decreases in collective diversity) and whether plural critics and gain policies mitigate it (Doshi and Hauser, 2024). Preregistration should include hypotheses, ablations (remove critic; freeze gain; swap policies), stopping rules, and telemetry (candidate sets, critic scores, reward-prediction error, temperature trajectory) to support external audit.

GS-3 is a proposal to turn a functional analogy into a testable engineering target. Its value hinges on rigorous comparisons to strong baselines, preregistered metrics and ablations, and transparent reporting. If its predictions fail, that outcome is informative—favoring alternative accounts such as scaffolded reflectivity or purely external regulation. If they succeed, they mark a step toward artificial systems that manage the tension between novelty, usefulness, and diversity by learning to regulate their own creative process (Holtzman et al., 2020; Chu et al., 2024; Izacard and Grave, 2021; Chen and Chan, 2024; Doshi and Hauser, 2024).