Curriculum engineering, as defined in Towards AI-Native Software Engineering (SE 3.0): A Vision and a Challenge Roadmap, is the systematic design, development and continuous refinement of curriculums that contain curated, organised and high-quality domain-specific knowledge, used as a more efficient alternative to training foundation models on vast amounts of unstructured, internet-scale data. It values high-quality data curation coupled with synthetic data generation over what the paper calls blind scraping of internet-scale data. A curriculum does not have to be designed by hand from top to bottom: humans can use AI to kick-start it, revise it manually, and, once it is stable, use AI again to generate synthetic data such as concrete examples that enrich it.
Usage
The term appears in the paper’s vision of Software Engineering 3.0 (SE 3.0), where the model behind its code synthesiser, FM.next, is expected to be trained on a high-quality software engineering curriculum — for example one inspired by the IEEE Software Engineering Body of Knowledge (SWEBOK) — so that it shows uniform competence across requirements reasoning, architectural design, implementation, testing, debugging and maintenance. The paper argues that a curriculum externalises knowledge from the model itself, so that it can be reviewed, reorganised or augmented in response to how the model performs, using observability data about where the model excels or underperforms. On this view curriculums, rather than the models, represent the intellectual property, and they need to be asset-managed and versioned so that they can be expanded and reused across models and use cases.
When It Applies
The paper offers a reference recipe for designing and maintaining a curriculum, drawing on IBM’s InstructLab as a prominent example of a systematic approach. Design starts by defining objectives and scope, identifying key domains and subdomains, and having domain experts outline core concepts, tasks and expected input-output specifications. The curriculum is usually structured as a hierarchical taxonomy, each node a task or knowledge area with examples, templates and evaluation rules; InstructLab suggests branching the root into knowledge, foundational skills and composition skills. Because the few manually curated samples at the leaves are often not enough for satisfactory instruction following, taxonomy-driven synthetic data generation — for instance with a teacher model — fills the gap. The curriculum is then tested for internal consistency and refined through pilot testing and community contributions, with a data flywheel recommended so that curriculum engineers change it to address problems observability data reveals, such as under-represented areas, emerging trends in the data and concept drift.
It assumes that observability data about the model’s performance is available to drive refinement, and the paper itself treats designing an effective software engineering curriculum as an open challenge. How well established the practice is: it is presented as one research group’s proposed strategy within a vision paper, supported by reference to recent systematic approaches such as InstructLab rather than by a measured result of its own.
Related Terms
- Software Engineering 3.0 (SE 3.0) — the proposed era whose models this approach is meant to train