Generative models of living development: an interaction-centered proposal and its implications for artificial life research

Artificial life has long pursued a double ambition: building systems whose organization reveals what life and agency actually require, and using those systems to clarify the conditions under which complex adaptive behavior arises. Cellular automata, evolutionary simulations, embodied robots, world models, and neural agents have demonstrated that sophisticated behavior can emerge from simple rules, recurrent coupling, and self-organization. Embodied and enactive approaches further established morphology, sensorimotor coupling, and organism-environment organization as explanatory resources rather than implementation details. Yet many artificial systems enter evaluation as already-formed controllers facing externally specified tasks. Their competence may be substantial, but the developmental organization through which events acquire significance remains outside the model. A new theoretical proposal published in Frontiers in Systems Biology argues that this omission is precisely where the most important questions about agency, development, and artificial life now lie.

The article, authored by Bin Yin of Fujian Normal University and Xiao-Rui Wu of Fuzhou University of International Studies and Trade, advances an interaction-centered theory that links computational generative models with experimental ecologies. In the computational sense, a generative model specifies how hidden states and causes give rise to observations and how beliefs and action policies are inferred. In the experimental sense, a developmental ecology generates the histories and observations from which causal and computational models are estimated. The authors contend that these two senses are methodologically coupled but not identical, and that treating the experimental arrangement as part of the explanation raises the question of how developmental meaning is constituted. Neither temporal extension nor physical stimulation alone determines what an event signifies for a living system. System-relative significance arises when an implemented mapping makes an event informative about conditions that matter to the system’s organization.

Code biology supplies one candidate mechanism for such mappings. An organic code consists of correspondences between entities in two independent domains, implemented by a third class of entities called adaptors. In this technical context, conventional means that the correspondence is not fixed by direct physical necessity and that alternative mappings are physically possible; it does not imply conscious agreement. The genetic code is the paradigmatic case because adaptor molecules implement correspondences between nucleotide triplets and amino acids. Whether comparable codes operate at neural or psychological levels remains an empirical question. The code-mediated concern hypothesis treats these mappings as testable mechanisms: a developmental event acquires system-relative functional significance when a maintained mediator maps a cue or relational pattern onto consequences for one or more concerns, thereby altering evidence, precision, habitual disposition, or action-policy selection. The resulting account locates functional significance in the relation among cue, mapping, and concern rather than in the cue alone, and yields a direct empirical requirement: changing the mapping should change the cue’s functional significance while leaving basic sensory discrimination intact.

The framework distinguishes reward and prior preference from concern, normativity, habit, and action policy. In reinforcement learning, reward is a formally defined signal used to update action values or optimize return; the term alone does not indicate whether the signal is generated by, or integrated with, the agent’s maintenance dynamics. Active inference already offers a richer formalization through prior preferences, uncertainty, and expected free energy. Concern addresses the constitutive and developmental question of what gives these quantities normative force; it does not replace them. A concern variable denotes a system-relative dimension of regulatory significance together with the implemented relations that connect observations and actions to it. The construct directs attention to how an organized vulnerability gives a preference normative force, how cues become evidence about that vulnerability, how those mappings are conserved or revised during development, and how stable reorganization can be distinguished from transient response or externally imposed valuation. In active-inference models, concern-related effects may be distributed across prior preferences, likelihood and transition models, habitual priors, precision, and hierarchical state representations.

Within this vocabulary, policy retains a restricted technical meaning. An action policy is a candidate sequence or conditional organization of actions evaluated under a model. Extending the term to habits, norms, joint commitments, tool-mediated capacities, and symbolic practices obscures important differences in organization and invites confusion with governance or public policy. A habit is a history-dependent disposition that makes some actions or action sequences more available, probable, fluent, or resistant to reconsideration. Active-inference models can represent habits as learned priors over policies generated through repeated action, and can examine how habitual and goal-directed control interact. In a broader biological sense, habits include sedimented sensorimotor routines, affective expectations, postural tendencies, and relational styles. They can bias formal policy selection without being identical to a selected policy, and they may organize action before an explicit choice set is represented. A history of reliable support may sediment as flexible approach and rapid recovery, whereas unpredictable threat may sediment as vigilance, hesitation, or rigid avoidance across contexts. Describing such trajectories only as newly selected policies would obscure their embodied and historical organization.

These distinctions are organized by the Layered Model of Agency, developed within cognitive semiotics by Juan Mendoza-Collazos and Jordan Zlatev. The model differentiates operative, reflective, joint, material-mediated, sign-mediated, and symbolic agency. The first three constitute original agency, in which action does not constitutively depend on artifacts; the latter three constitute enhanced agency, in which material and semiotic resources expand the capacity to act. Fundierung denotes an asymmetric relation in which higher layers depend on and reorganize lower capacities. Operative agency comprises embodied, often pre-reflective action in which skill and intention are expressed in the performance itself. Reflective agency introduces explicit anticipation, evaluation, or commitment. Joint agency arises when agents coordinate in relation to a shared or mutually regulated activity. Material-mediated agency occurs when artifacts become constitutive resources for action. Sign-mediated agency recruits nonlinguistic signs to guide action across absent, counterfactual, or socially shared situations. Symbolic agency recruits linguistic and other symbolic systems that support abstraction, explicit norms, narratives, institutions, and reasons. The layers do not form a scale of moral worth, and a single situated activity may recruit several at once.

The model also places subjectivity at an evidential boundary. Within its cognitive-semiotic formulation, genuine agency entails at least minimal subjectivity. Behavioral performance alone cannot establish genuine agency. Three levels of attribution help preserve this distinction: agentic functionality denotes autonomous coordination of sensing and action relative to internal variables or goals; agency-like organization adds developmental integration with self-maintenance, normativity, historical individuality, and reciprocal coupling; phenomenal or genuine agency requires adequate evidence of a first-person perspective or sense of agency. Synthetic systems can demonstrate the first two. The third remains an open empirical and philosophical question, especially for artificial systems. This distinction supports a substantive artificial-life program: a system that preserves itself, acquires concern-sensitive mappings, forms habits, coordinates with others, appropriates artifacts, and uses signs productively would constitute a substantial theoretical result even if subjectivity remained unresolved. Fluent behavior, ungrounded linguistic self-report, or anthropomorphic description would remain insufficient grounds for inferring experience.

The Human-Rat Interaction Paradigm (HRIP) provides an empirical platform on which parts of this theory can be progressively tested. As a paradigm, HRIP organizes developmentally timed human-rat interaction histories and relates them to later emotional, cognitive, social, physiological, and neural trajectories. Its inaugural implementation enrolled 48 one-month-old male Sprague-Dawley rats and assigned 16 animals to each of three conditions: positive early-adolescent experience, negative early-adolescent experience, or ordinary laboratory control. Designated human caretakers delivered the three-week intervention, and interaction behavior stabilized by the second week. Later assessments sampled anxiety and exploration, adaptation to novelty, learning and memory, responses to toys and unfamiliar rats, behavior in trapped-peer and food-access tasks, and social competition. Because the published report included no a priori power calculation, group contrasts should be treated as preliminary estimates rather than precise effect-size benchmarks. The male-only sample further limits inference to males; generalization across sexes requires replication in adequately powered, sex-inclusive designs with planned sex-by-history tests. HRIP itself remains open to female, male, and mixed-sex implementations and to comparisons across developmental periods.

Six propositions and eight predictions define tests of the theory. The propositions concern organizational normativity, code-mediated significance, historical sedimentation, ecological generation, layered emergence, and perspective-matched evidence. The predictions specify discriminating designs: nested model comparison fitting reward-only, standard active-inference, habit-augmented, and code-adaptor models to the same longitudinal data; cue remapping and adaptor perturbation; habit-policy dissociation; reciprocity and emergent joint organization through contingent-versus-yoked comparisons; material and sign mediation; multimodal latent trajectories; developmental timing and sex moderation; and layer-specific evidence. A code-adaptor model is supported only if it improves out-of-sample prediction specifically under cue remapping, partner change, or cross-context transfer. Better fit to training data is insufficient. If standard active inference with preferences, transitions, and precision predicts intervention and transfer effects as well as the code model, the code construct adds no demonstrable explanatory value. In that case, concern may still guide constitutive interpretation, but the mechanistic code claim should be weakened.

For artificial life, the program specifies an ecology alongside an objective function. The design identifies what the system maintains, which partners and affordances are available, how interactions are contingent, which artifacts and signs can be appropriated, and how plasticity changes over developmental time. Counterfactual histories are compared within a shared architecture, while alternative architectures encounter matched histories. This factorial logic separates organizational effects from cumulative exposure. Contemporary generative systems can display impressive interactive competence without acquiring it through their own developmental histories; competence inherited from human data, prompts, or design scaffolds should not be treated as evidence of developmentally grounded agency. The authors argue that evaluation should distinguish capacities constituted through the system’s own embodied and relational history from those that depend on human semantic infrastructure. Across biological and artificial systems, the framework directs inquiry toward how operative, joint, mediated, and symbolic forms of agency emerge, how interventions alter them, and what evidence warrants each attribution. Functional organization and phenomenal agency remain distinct explanatory targets.

Subject of Research: Generative models of living development: an interaction-centered proposal and its implications for artificial life research

Article Title: Generative models of living development: an interaction-centered proposal and its implications for artificial life research

Article References: Yin, B., & Wu, X.-R. (2026). Generative models of living development: an interaction-centered proposal and its implications for artificial life research. Frontiers in Systems Biology, 6, Article 1904510. https://doi.org/10.3389/fsysb.2026.1904510

Image Credits: AI Generated

DOI: 10.3389/fsysb.2026.1904510

Keywords: Generative, models, living, development, interaction-centered, proposal, implications, artificial, life, research, scientific research