self-scaffolding

Self-scaffolding is a way for an AI model to set up the scaffolding it needs before solving a problem, then solve the problem on that scaffolding. Just as a construction worker assembles scaffolding before building, the model decides for itself which tools to use and in what order before it starts working. In the past, humans wrote the manuals and grading criteria in advance and handed them to the model — but in this approach, the model creates that manual on its own.

Taken a step further, the model doesn't just build the scaffolding; it also generates the problems to practice on and the solutions to them, feeding that material back into its own training in a loop. In effect, it identifies its own weak spots and keeps generating practice problems tailored to them. This reduces the burden of humans having to hand-craft problem sets, but it also means the model's performance now depends on how accurately it can evaluate and improve itself.

This concept has drawn attention because it enables a self-improvement loop that requires less human involvement. The underlying expectation is that a model finding and filling its own gaps can improve faster and more broadly than traditional training methods that rely on human-made problem sets and grading tools.

News coverage describes it as "a method where the model builds for itself the tools and procedures — the scaffold — needed to solve a problem, then solves the problem on top of it." The concept was first introduced by the Ornith development team in their earlier model, Ornith-1.0, and the successor, Ornith-1.5, took it further into a self-contained loop where the model also generates the tasks to solve and the grading criteria itself. It's worth noting that self-scaffolding refers to a training and inference approach, not the name of a specific model.