Toward Agent-based Educational Science: Rethinking Educational Research in the Age of AI

Synthesis: Zhang, Jiang and Tang argue that educational science suffers from a structural mismatch: pedagogical and AI-driven learning innovations are deployed faster than classroom-based empirical research can evaluate their developmental impact, so evidence arrives only after large-scale implementation and the field accumulates descriptive findings rather than mechanistic, testable theory. Their proposal is agent-based educational science (AES), a paradigm in which educational theories are formalized as interacting agents embedded in structured environments, so that developmental processes that are slow, risky or ethically infeasible to test in classrooms can be explored in silico. The concrete instantiation they introduce is Student Development Agents, three-layer models combining a learner model of cognitive, motivational and socio-emotional variables, an environment model of pedagogical, social and technological conditions, and an LLM-driven simulation engine that iterates behavior generation and state updating to produce longitudinal developmental trajectories. The authors explicitly do not claim that Simulation replaces empirical research: classrooms remain the site of grounding and ethical accountability, and empirical data are recast as calibration, validation and boundary conditions. The paper is a position preprint. It reports no empirical evaluation of AES or the Student Development Agent platform, and its own simulated learner literature review concedes that LLMs still miss inter-individual variability and that validation of generative social simulation remains the field's central unresolved challenge.

Key Findings

1. The diagnosis is a structural mismatch, not a shortage of studies. Educational innovation outpaces verification because the empirical research paradigm of classroom study is slow, context-bound and predominantly descriptive, and the authors complement the resulting evidence-innovation gap with the field's documented replication crisis.
2. AES is defined by epistemic logic, not by a particular technology. The paradigm formalizes educational theories as interacting agents embedded in structured environments, making Simulation a core epistemic tool for generating developmental trajectories and exploring counterfactual educational designs rather than a downstream add-on to empirical work.
3. Student Development Agents are offered as the minimal instantiation. An SDA is a three-layer architecture: a learner model parameterized by cognitive, motivational and socio-emotional variables, an environment model encoding pedagogical, social and technological conditions, and a simulation engine that runs LLM-based inference in an iterative cycle of behavior generation and state updating.
4. The evidence-to-simulation loop is LLM-centred. Heterogeneous educational knowledge, including meta-analytic results, intervention datasets, interviews and qualitative narratives, is translated into simulation-ready constraints through LLM-based reasoning, retrieval-augmented generation and fine-tuning, with prompt and context engineering used to run what-if scenarios across learner profiles.
5. Empirical data keep a central but reconfigured role. Instead of being solely the site of discovery, empirical studies provide calibration, validation and boundary conditions for agent models, and discrepancies between simulated and observed trajectories are treated as informative signals about underspecified theory or violated assumptions.
6. The paradigm is argued by cross-disciplinary analogy. The paper situates AES alongside the computational turns in economics, sociology and political science, and alongside climate science and systems biology, with the closest parallel being medicine's move from clinical trials to in-silico drug testing for accelerated discovery at reduced ethical risk.
7. Ethics and governance are treated as design constraints, not afterthoughts. The authors require privacy-by-design with local differential privacy, secure multi-party computation and federated learning, learner-centered rather than surveillance-driven digital twins, algorithmic accountability extended to in silico experiments, and open internationally governed SDA repositories that address data sovereignty.
8. No empirical evaluation is reported. The paper is a pre-print position argument: Student Development Agents are described conceptually, with the operational architecture and evidence-integration mechanism specified but no simulation results, validation study or classroom comparison presented.

The structural mismatch in educational science

The paper opens from a diagnosis rather than a gap in the literature. For decades the field has relied on empirically grounded but resource-intensive and ethically constrained studies of how teaching methods and learning environments shape cognitive and socio-emotional development, and those studies have produced indispensable insights. The problem is tempo. Adaptive learning systems and AI-driven applications arrive faster than the evaluation apparatus can assess them, so evidence is generated after large-scale implementation has already occurred. The authors compress this into the claim that education innovates faster than it can verify its impact, and note that educational theories consequently often remain static narratives rather than mechanistic models able to generate predictions across contexts and scales.

A second strand of the diagnosis is developmental. Education's scientific object is characterized as trajectories, not events: learning outcomes are shaped over time by the interaction of cognitive, motivational, socio-emotional and contextual processes, yet most empirical studies capture narrow temporal windows and short-term performance measures. The authors call this constraint epistemic as much as logistical, because when theories concern processes unfolding over months or years while methods observe only fragments, theoretical progress becomes incremental and hard to cumulate. Existing quantitative and qualitative strands are presented as fragmented, and the traditional linear paradigm is contrasted with an agent-based paradigm that integrates theory building and empirical validation into one iterative loop.

What agent-based educational science proposes

AES is deliberately not defined by its technology. The authors frame it as a shift in epistemic logic: rather than relying exclusively on post hoc observation of implemented interventions, it formalizes educational theories as interacting agents embedded within structured environments, with agents encoding theoretically grounded representations of Learners, educators and contextual forces, and environments specifying pedagogical, social and technological conditions. The resulting Simulation generates developmental trajectories, explores counterfactuals, and constrains theoretical claims under conditions that are slow, costly or ethically infeasible to test directly. Precedent is drawn from computational economics, sociology and political science, and from climate science and systems biology, all of which integrated large-scale simulation as data complexity and ethical sensitivity grew; medicine's shift from clinical trials to in-silico testing supplies the strongest analogy.

Student Development Agents as the minimal instantiation

To make the paradigm operational, the authors introduce Student Development Agents as a minimal instantiation rather than a standalone platform or predictive oracle. An SDA is defined by its capacity to generate developmental trajectories instead of isolated behavioral responses, and its architecture has three interacting layers: a learner model parameterized by cognitive, motivational and socio-emotional variables representing dynamic internal states; an environment model encoding pedagogical, social and technological contexts, including AIED tools and teaching strategies; and a simulation engine that uses LLM-based inference to execute an iterative cycle of behavior generation and state updating, feeding outcomes back into the learner model so that static profiles become evolving predictions.

It is worth being precise about what kind of agent this is. In this paper the agents primarily stand in for learners and teachers rather than acting as autonomous scientific agents: multiple SDAs can interact with one another and with simulated teacher, parent or peer agents to reproduce the social ecosystems of real settings, and the research instrument is the simulation apparatus that orchestrates them, not the agent as a decision-maker. The environment model operates in two modes, either holding a real-world context static as a controlled baseline or evolving dynamically alongside the agents as a self-contained digital-twin sandbox. The authors state that SDAs are not intended to model or predict the futures of specific learners; their purpose is to explore classes of developmental possibilities under specified conditions, which they present as a guard against misuse in high-stakes contexts.

Simulation, theory development, and existing research methods

The paper positions AES as continuous with, not opposed to, prior student simulation research. It distinguishes individual-level modeling, where representational alignment between human and artificial cognitive representations is the target, from population-level simulation, the LLM-powered agent-based modeling used in the social sciences to construct large-scale citizen societies. Education is read as already benefiting from LLM-based simulated students built through prompt-based parameterization of personality, knowledge state, motivation and self-regulated learning behavior, superseding earlier symbolic and parametric approaches such as SimStudent, alongside fine-tuning on classroom dialogue for behavioral fidelity. The stated gap is developmental: most simulation studies remain static, reproducing momentary responses or short-term behavior, and student development has not been a central concern even as self-evolving agents with long-term memory and dynamic state tracking begin to appear.

That gap is where the theory development claim sits. Computational models let researchers formalize theories, explore counterfactual scenarios and generate predictions that can be selectively probed against data, which the authors argue supports the cumulative, falsifiable progression that mature learning science is said to lack. AES is therefore offered as an extension of research methods rather than a replacement for them, and it sits in the same family as existing work in learner modeling, learning analytics and intelligent tutoring, from which it differs by treating time-extended development as the modeled object rather than momentary performance. The authors also tie the paradigm to developmental theorizing across cognitive, motivational and socio-emotional domains that a single short study cannot observe at once.

Evidence, replication, and the ethics of simulated learners

The most consequential claim the paper makes about evidence is a redistribution of labor. Simulation does not replace observation; it reshapes the division of labor between theory generation and empirical validation. Simulation becomes the space for systematic exploration and hypothesis generation, while empirical studies supply grounding, correction and ethical accountability, so that designs are informed by predictive, theory-driven exploration before implementation rather than only evaluated after it. Discrepancies between predicted and observed trajectories are treated as informative rather than as failures, indicating where theories are underspecified, assumptions violated or contextual factors overlooked. The authors present this as a route to a predictive science of learning that is also ethically safer, since risky or infeasible conditions can be tested in silico, and the analogy with replication is implicit: a formalized model is a claim that can be re-run, varied and falsified in a way a static narrative cannot.

The ethical program is correspondingly detailed. Building digital twins of learners requires high-resolution longitudinal data, and the authors argue that governance must evolve alongside methodological innovation. They call for privacy-by-design embedded from the outset through local differential privacy, secure multi-party computation and federated learning architectures, with data minimization and purpose limitation; for digital representations that remain learner-centered rather than surveillance-driven, so they are not used for behavioral manipulation or automated decisions about learners' futures, with mandatory transparency and rights to contest and opt out; and for algorithmic accountability and bias audits to extend beyond empirical AI systems to in silico experiments, with human oversight in interpretation. The institutional horizon is a set of open, internationally governed SDA repositories, which the authors argue must be established with robust AI Governance addressing data sovereignty before systems are deployed at scale.

What this means for practice

• Researchers. Formalize a theory you already test in classrooms as an interacting-agent model before running the next study, and read a mismatch between simulated and observed trajectories as a signal about underspecified theory or violated assumptions rather than as a failed experiment. The paper assigns simulation the hypothesis-generation role and empirical studies the roles of grounding, calibration and boundary conditions.
• Researchers. Use Student Development Agents to probe conditions that are slow, costly or ethically infeasible to test directly, such as long-horizon development or risky interventions, and reserve classroom study for confirming what the simulation proposes.
• Administrators. Make privacy-by-design a condition of approval before any learner digital twin is built from institutional data, requiring local differential privacy, secure multi-party computation or federated learning together with data minimization and purpose limitation.
• Administrators. Fund and govern shared model repositories rather than institution-specific ones, and extend algorithmic accountability and bias audits to in silico experiments, not only to deployed AI systems; the authors treat open, internationally governed repositories with data-sovereignty agreements as a precondition for deploying the paradigm at scale.

Limitations

• The paper is a pre-print position piece, and it reports no empirical evaluation: there is no simulation study, no validation of Student Development Agent trajectories against real student data, and no test of the claim that in-silico experimentation closes the evidence-innovation gap.
• Student Development Agents are introduced as a minimal instantiation whose architecture and evidence-integration loop are specified conceptually, and the authors present the vision as requiring a unified global effort from psychologists, computational linguists, data scientists and educators rather than as something already demonstrated.
• The authors' own review of representational alignment concedes that current neural representations fail to capture high-level semantic and abstract dimensions of human cognition, and that while LLMs show aggregate-level alignment with human judgments they remain limited in capturing inter-individual cognitive variability and the motivational, socio-emotional and ethical dimensions that shape real decision-making; most existing simulation work is static, and developmental trajectories remain largely overlooked.
• Validation is named in the cited literature as the central challenge for generative social simulation, a caveat the paper does not resolve, and the Simulation-versus-classroom boundary rests on a distinction the paper asserts rather than tests: it offers no account of how much predictive authority simulated trajectories should carry in decisions about real programs, or how the digital-twin data needed to build them could be assembled under the privacy regime it simultaneously demands.

Connected Concepts

• Agentic AI — the LLM-driven agent architectures that make the paradigm practically achievable for the first time
• Simulation — the core epistemic tool of AES, generating trajectories and counterfactual educational designs in silico
• Simulating Students — the closest existing research family, which the paper seeks to move beyond static, momentary simulation
• Learner Modeling and Adaptive Instruction — the learner model layer, parameterized across cognitive, motivational and socio-emotional variables
• Theory Development in AI in Education — the cumulative, falsifiable theory building the authors argue education lacks
• Research Methods in AIED — the slow, fragmented classroom paradigm AES is offered as a complement to rather than a replacement for
• Learning Sciences — the explanatory depth the authors aim to fuse with the generative capacity of LLM agents
• Large Language Models (LLMs) — the inference and knowledge-synthesis engine of the Student Development Agent loop
• Intelligent Tutoring — an adjacent modeling tradition whose simulated students the paper treats as predecessors
• Learning Analytics — existing data-driven educational modeling whose temporal scope AES explicitly widens
• Privacy — privacy-by-design, federated learning and data minimization as preconditions of learner digital twins
• AI Governance — internationally governed SDA repositories and accountability extended to in silico experiments
• Ethics — bias audits, transparency, contestation rights and human oversight in simulation-based research

Connected Articles

• Simulating Students with Large Language Models: A Review of Architecture, Mechanisms, and Role Modelling in Education with Generative AI — Systematic review of LLM-based simulated students, the literature this position paper builds on
• Towards Valid Student Simulation with Large Language Models — Validation concerns in LLM student simulation, the challenge AES inherits
• INSIDE the Student's Mind: Jointly Modeling Latent Reasoning and Action in LLM Student Simulators — What happens inside LLM student simulators when they reason
• Who Am I? History-Aware Profiles for Student Simulation in Tutoring Dialogues — Memory and history in student simulation, the developmental direction AES calls for
• CogEvolution: A Human-like Generative Educational Agent to Simulate Student's Cognitive Evolution — Simulating the evolution of student cognition over time
• StudentSim: Training LLM-based Student Simulators — LLM student simulators as research instruments
• From Precision Medicine to Precision Education: A Vision for AI-Powered Student Digital Twins, Preventive Student Success, and Career-Aligned Academic Pathways — Student digital twins, the construction AES treats as ethically loaded
• AgentSchool: An LLM-Powered Multi-Agent Simulation for Education — Multi-agent simulation of educational dynamics at population level
• Modelling Individual Participants as LLM Agents in Collaborative Problem Solving Simulations — LLM agents used to simulate collaborative learning environments

Citation

Zhang, Y., Jiang, J., & Tang, X. (2026). Toward agent-based educational science: Rethinking educational research in the age of AI. Educational Psychology Review, 38(1), 113.