Abstract

In an era where abundant high-quality open educational resources are readily accessible online, the potential for personalized learning through adaptive, dynamic curricula has yet to be fully realized. While a few experienced users may manage to combine diverse sources to create customized learning paths for themselves, most learners are left navigating pre-packaged, often suboptimal learning modules. This paper explores the potential of artificial intelligence to bridge this gap by dynamically generating custom learning paths from OERs, leveraging large language models and advanced data modeling to retrofit existing educational ontologies. We introduce a framework for creating adaptive curricula that are dynamically sequenced based on user needs, behaviors, and learning goals, thereby unlocking new opportunities for personalized education and making OERs accessible and effective for a broader audience.

Introduction: The Problem

High-quality, open-access learning resources are widely available online, from self-contained MOOCs like edX to multipurpose platforms such as YouTube. An exceptionally savvy and diligent user would be able to combine coursework and sections of courses from different platforms to create a personalized learning path through a dynamic heterogeneous curriculum that suits their background, aptitudes, and learning goals. Our experience shows that using this kind of "Frankenstein curriculum" leads to better learning outcomes and engagement compared to predefined curricula on a single platform or even adaptive coursework in the context of an edtech application. This is not unexpected because pre-packaged curricula designed with a given set of learner profiles in mind are unlikely to be optimal for learners whose characteristics, time commitment and expectations differ significantly from the archetype (Badali et al., 2022). However, only a small minority of potential learners are knowledgeable and experienced enough to proactively adapt open coursework to their needs. A general-purpose solution that could create and dynamically update custom learning trajectories personalized to individual learners’ needs would unlock the latent potential of OER ecosystems.

Retrofitting OER Ontologies Using LLMs

The idea of making online learning resources more accessible and combinable is not a new one. A key aspect of the project of making such resources more discoverable, interoperable and modular is the development of ontologies, such as OER Schema (https://oerschema.org/), that attempt to enhance the way open courseware is located, authored, shared, and remixed, with the goal of making OERs more interoperable and accessible to machines. However, major OER platforms have not actively adopted such standards, and a vast domain of less formal learning resources, such as YouTube videos, are generated organically, by a large number of individuals or entities and, as a result, do not lend themselves to centralized ontology modeling. In addition, platforms like Coursera that aim to offer a comprehensive curricular experience do not have an incentive to break their coursework down into highly granular units that can be recombined by third parties.

Discoverability, Interoperability and Modularity

Discoverability is an essential characteristic of online learning resources because it can maximize their reach, effectiveness, and adaptability. It allows students, educators, and self-learners to efficiently locate high-quality resources that meet specific learning objectives and preferences. When educational resources are easily searchable and accessible, they contribute to more equitable access to knowledge, bridging gaps for learners who may not have traditional educational opportunities. Discoverable content also allows educators to integrate diverse resources into their curricula, enriching the learning experience through a variety of perspectives and instructional formats.

Interoperability is equally vital for fostering a flexible, learner-centered educational environment. Interoperability enables different platforms, tools, and resources to communicate and function seamlessly together, which not only enhances usability but also empowers educators and learners to use the resources within their own technological ecosystems. This compatibility facilitates seamless transitions between platforms, allows for efficient data sharing, and encourages collaborative, interdisciplinary learning experiences.

Modularity further enhances this by allowing courseware and learning materials to be organized in relatively free-standing units or components, increasing opportunities for creating custom combinations. This modularity allows educators to customize lessons to suit specific learner needs, adapt resources to various contexts, and update or reorganize content without disrupting the whole.

Sub-Modular OER Components

Taking the idea of creating modular learning resources to its extreme implies courseware designed as combinations of relatively small learning modules. These would be designed to follow from or lead to other modules in a way that maximizes possible combinations and chaining opportunities. Learning modules would be as atomic as possible, rather than being embedded in monolithic combinations as a succession of strictly sequenced chapters. To take the atomic metaphor a bit further, we could portray larger learning modules as highly varied sets of molecules assembled from a common set of atomic learning units, and curricula as large, organic molecules that chain strings of atoms together. Ideally, these large molecules can be transformed and adapted for specific applications. From the perspective of typical online courseware, where modules are relatively monolithic, this approach would require mining courses for sub-modular components at the level of lessons or chapters.

Post-Factum Modeling with AI

LLMs and multimodal AI make it possible to process OERs and model them in terms of a given ontology. They could also be used to break them down into sub-modular components, their atomic units, and tagging them with metadata that would provide the information needed to sequence them with modules from other platforms or sources. This post-factum ontologization and tagging would make it possible to construct dynamic heterogeneous curricula. In fact, the viable sequences of modules could form tree or graph structures where atomic learning modules are nodes and a learner's possible transitions from one module to the other are branches or edges. Using an LLM, it would also be possible to automatically generate assessment modules associated with each edge of such a graph. A learner would traverse the graph based on choices resulting from the outcomes of these assessments, combined with the learners’ characteristics and preferences. The result would be learning paths that are optimized for each learner and that can dynamically adapt to and incorporate new OERs, coursework, video tutorials and the like as they become available.

Dynamic Learning Path Generation

Below, we propose a strategy for dynamically generating adaptive curricula across platforms using AI models. For the moment, we ignore some constraints such as paywalls, lack of API access or legal issues that may prevent the full implementation of this strategy across all online learning resources. However, the universe of such resources is large enough that even working with only the highly accessible subset of resources can be very effective.

Map OERs

A necessary first step in this process is cataloging online learning resources and breaking them down into submodular components as described in the section on post-factum modeling above. This can be done by leveraging search engines and LLMs that have already crawled and trained on these resources, or it can involve a proprietary crawler, possibly in combination with search engines and mainstream AI models. In any case, it is important that the resulting graph (where learning modules are nodes and a learner’s transition from one to the other are edges) be dynamic and constantly updated because of the evolving landscape of OERs and online learning resources in general. The experiences of learners using the resulting curricula can be used to fine tune the graph with data that allows for more optimal traversals through it, ones that are likely to lead to higher completion rates and better learning outcomes.

Create Dynamic Assessment Tools

A key decision point in the process of creating optimal curricula for learners involves picking the next module based on a learner's experience with (and learning outcomes from) a previously completed module. OERs often have built-in assessments that can provide valuable information about the learner’s experience and readiness to move on to a given next module. But these are not sufficient for our purposes for a number of reasons. The results of these assessments may not be machine-accessible; There may not be interim assessments at the sub-modular level which our mapping process would have extracted as atomic components; and many learning resources such as video tutorials may not have associated assessments. The assessment tools can be dynamic and interactive, consisting of a conversation with an LLM prompted to determine the level of mastery of a topic by generating questions and mini-quizzes on the fly, taking as input the content and target learning outcomes of a completed module.

Get to know the Learner

The learner’s behavior, choices, successes and failures as they progress through a learning path provide key information for generating and adapting the continuation of their path. However, it is also important to get, directly from the user, information about their expectations and aspirations, their time commitment, as well as their background and capabilities, possibly using an initial assessment generated dynamically once the basic learner data is gathered. Clearly, this process of getting to know the learner, which will then dictate key aspects of curriculum generation, can be done effectively using a conversational LLM as an open education guide. This guide would accompany the learner through their trajectory across the curriculum and can both inform them as well as solicit information from them on a continuous basis. The open education guide can also play a useful role in compensating for some rough transitions as the dynamic curriculum switches context between different platforms that are not seamlessly integrated.

Generate and Update Learning Paths

The process of generating a learning path or curriculum is a fairly mechanical one, based on the mapping of online learning resources and the resulting cross-platform graph as well as the dynamic assessment tools associated with the transitions between modules in the graph, and the input gathered from the learner initially and on a continuous basis over the course of their learning trajectory. The challenge for this component is around UI/UX, requiring well designed learning management and user management functionality.

Optimize Path Generation with Reinforcement Learning

In addition to using mainstream AI models to ontologize learning resources, to generate assessment tools and to get to know the user, the dynamic heterogeneous curriculum strategy involves training a specialized model to optimize the performance of the open education guide. We assume that learners with similar characteristics, goals and previous learning trajectories would have a similar likelihood of success in completing and benefiting from specific subsequent learning modules. This means that training on this type of learner data in order to recommend an optimal next step could be productive at scale. However, this machine learning strategy would need to be embedded in a model informed by a specific theory of learning and a pedagogical strategy. This is necessary for a number of reasons, including the following. New learning resources become available all the time and need to be given relative priority in order to fine tune the model for them; The model’s initial training can be biased towards a specific majority of early users, therefore providing suboptimal recommendations to a monority of learners who would join later; And until the scale of training data is sufficiently large, the model would not otherwise be able to rely on the heuristics provided by a pedagogical strategy.

Expected Evolution of OERs

The proposed initiative to create dynamic heterogeneous curricula is meant to work as seamlessly as possible with existing open educational resource providers, offering an innovative approach to personalized learning. The core idea is to break down existing OER coursework into atomic modules or components that can be flexibly combined based on a learner’s unique needs. These small, self-contained units would come pre-tagged with relevant characteristics, such as difficulty level, content focus, and skill type. As mentioned above, this tagging system would allow for easy recombination and sequencing to align with individual learning objectives. Faced with this post-factum processing of their content, OER publishers would be compelled to organize their content around such modular components, simplifying integration across diverse educational platforms in a way that they can control and contribute to. This means that the dynamics of creating Frankenstein curricula has the characteristic of built-in obsolescence, and would be replaced by a more structured process over time.

As the educational landscape continues to evolve, we are seeing a shift toward AI-powered educational tools, with many OERs likely to incorporate advanced tutoring capabilities. Such AI-powered tutors would guide learners through courses, adapting content in real-time to optimize comprehension and retention. This transition to AI-driven education would further enhance the modular approach, as the AI could select and sequence sub-modular OER components dynamically, based on continuous assessment of learner progress within a given platform. AI integration would support real-time feedback, making educational resources more responsive to students’ needs and reducing reliance on traditional assessment methods. However, given that users are likely to interact with difference platforms during the course of their learning experiences, an AI guide that is persistent across platforms would be useful. It is interesting to speculate whether and how the platform specific AIs and the cross-platform ones described above would coexist and possibly collaborate.

Ultimately, the role of an open education guide, as a tool to help users navigate OERs, would evolve. Initially, the guide would focus on cataloging, tagging, and assessing resources. Over time, as AI systems become more adept at managing educational content and customizing learning experiences, the guide's function would shift. Rather than cataloging or assessing resources, its primary purpose would become streamlining the learner's journey across these modules. By dynamically advising on the user’s learning experiences, the guide would help maximize learning efficiency, allowing learners to achieve their goals more effectively, faster and with higher fidelity to personal goals.

Related Work

OATutor

The OATutor project (https://www.oatutor.io/) is an open-source adaptive tutoring system designed to advance research in intelligent tutoring systems and enhance learning experiences in mathematics and other subjects. It integrates advanced AI techniques to deliver personalized, interactive problem-solving support, leveraging natural language processing for generating hints and feedback. The system includes a curated library of Creative Commons-licensed content and supports research in learning sciences. By combining adaptability, accessibility, and open educational practices, OATutor aims to make high-quality tutoring scalable and impactful, with demonstrated efficacy in improving learning outcomes. OATutor provides a Creative Commons-licensed problem library and leverages Bayesian knowledge tracing for skill mastery, combined with built-in A/B testing capabilities. It is a platform for experimentation and analysis meant to facilitate innovation and replication of adaptive learning studies (Pardos et al., 2023).

TUMO Labs

Implemented by the TUMO Center for Creative Technologies (tumo.org), TUMO Labs is a technology education program for university-age students and young professionals based in Yerevan, Armenia (tumolabs.am). The program partners with technology companies to provide external R&D services, executed by students in a project-based-learning (PBL) environment under the guidance of industry and academic mentors. To address knowledge gaps required for specific PBL modules, students use open educational resources and other online tools for autonomous learning. TUMO Labs provides guides and assessments that implement this heterogeneous curriculum approach, serving as a coach and wayfinder in the online learning space. Currently, the catalog of resources, assessments, and guidance heuristics are developed manually by the TUMO Labs team. Despite this manual approach, the program has significantly improved course completion rates and learning outcomes (Rahimi et al., 2024). The introduction of an AI guide is expected to further enhance its scope and effectiveness.

Conclusion

The future of OERs lies in their capacity to adapt to each learner's unique needs, harnessing AI to generate dynamic, customized learning pathways. By decomposing educational resources into modular, recombinable components and utilizing LLMs to assess, sequence, and personalize curricula, we pave the way for a more flexible, learner-centered educational experience. The adaptive AI-based "Frankenstein curriculum" model represents a transformative shift in educational resource utilization, creating pathways that are not only responsive to individual learning goals but also scalable across diverse educational contexts. As AI-powered educational guides evolve, their role will shift from cataloging resources to optimizing the learning journey itself, ensuring that OERs reach their full potential in fostering meaningful, personalized, and accessible learning experiences across platforms.

References

Badali, M., et al. (2022). The role of motivation in MOOCs’ retention rates: A systematic review. Technology, Education, and Learning Research Policy (TEL-RP). Retrieved from https://telrp.springeropen.com/articles/10.1186/s41039-022-00181-3

Pardos, Z. A., et al. (2023). OATutor: An open-source adaptive tutoring system and curated content library for learning sciences research. ACM Digital Library. Retrieved from https://dl.acm.org/doi/fullHtml/10.1145/3544548.3581574

Rahimi, A. R. et al. (2024). A tri-phenomenon perspective to mitigate MOOCs' high dropout rates: The role of technical, pedagogical, and contextual factors. Smart Learning Environments Journal. Retrieved from https://slejournal.springeropen.com/articles/10.1186/s40561-024-00297-7