Research is often described as a journey from idea to publication. The moments that stay with me, however, are the ones that show an idea continuing to travel: into a wider conversation, into other people’s reading lists, and eventually into systems that can be tested in practice.
I am glad to share that BOLD and the EdTechnical podcast selected my AIED 2026 paper, From Untamed Black Box to Interpretable Pedagogical Orchestration: The Ensemble of Specialized LLMs Architecture for Adaptive Tutoring, as further reading for the episode “Why chatbots aren’t the future of educational AI”.
Published on 20 August 2026, the episode reports from the Festival of Learning in South Korea and asks a question that matters deeply to my work: how should we decide where generative AI belongs in learning, where it does not, and what must be designed around it?
It is important to describe the recognition precisely. I was not a guest on the episode, and BOLD did not independently review or endorse the paper’s findings. The paper was selected as further reading alongside the episode. That is still meaningful to me because the episode’s shift away from generic chatbots and toward fit-for-purpose educational systems closely matches the design principle behind ES-LLMs.
Why ES-LLMs goes beyond a chatbot
The ES-LLMs architecture separates pedagogical decision-making from language generation. A deterministic orchestrator coordinates specialised agents for tutoring, assessment, feedback, scaffolding, motivation and ethics. An interpretable learner model informs what the system should do next, while an LLM handles how that action is expressed in natural language.
This distinction matters. A fluent answer is not necessarily a good teaching decision. A general-purpose model can sound helpful while giving away an answer too early, overlooking a learner’s current state, or applying an instructional strategy inconsistently. ES-LLMs makes those decisions explicit and auditable. Rules such as “attempt before hint” and limits on assistance can be enforced, while agent traces and constraint checks can be inspected.
The paper reports evidence from expert review, multi-model evaluation and simulation. Those findings belong to the study and its stated conditions; they should not be treated as independent validation by BOLD. What the further-reading selection signals is something different: the problem ES-LLMs addresses is part of a wider international conversation about the future of educational AI.
From research architecture to working applications
The architecture is now powering the LearnAdapt family of AI applications. That transition from research contribution to implementation is especially satisfying. It creates a practical loop in which architectural ideas inform applications, while experience from those applications sharpens the next questions for research.
This does not mean that chat has no place in education. Conversation can be a valuable interface. The stronger claim is that the chat window should not be the whole educational system. The more important questions are whether the system has a coherent pedagogical purpose, whether educators can understand and shape its behaviour, whether learner state is represented responsibly, and whether safeguards are structural rather than merely prompted.
I am grateful to the AIED community for giving this work a scholarly home, and to BOLD and EdTechnical for including it in the wider discussion. I also appreciated meeting Bryan Richardson from the Gates Foundation during AIED 2026 in South Korea—one of the contributors featured in the episode. Encounters like these remind me that educational AI advances through conversations across research, practice, funding, design and public-interest work.
You can explore the work through the following links:
- BOLD / EdTechnical episode
- Published paper and DOI
- Free arXiv version
- Open-source research toolkit
- Live ES-LLMs sandbox
I welcome conversations with researchers, educators and builders who are interested in interpretable pedagogical orchestration, learner modelling, or the design of educational AI that can be examined rather than simply trusted.
Leave a Reply