Scaffold-mediated training lets models keep some learned skills after scaffolds vanish

Researchers have proposed scaffold-mediated post-training, a method that lets procedural scaffolds evolve alongside model parameters and then distills some of that behavior back into the model.

The paper argues that today’s post-training usually optimizes parameters while inference-time scaffolds are designed separately. That separation makes it harder for models to acquire complex strategies that depend on procedures, tools, or multi-step routines. The proposed approach organizes scaffolds into an evolvable graph and updates them through discovery, distillation, and recompilation.

In the authors’ Skill Training implementation, automatically discovered skills improved FeatureBench pass rate by 8.1 percentage points. After progressive distillation, the model retained 85.2 percent of the with-skill pass rate and reached a 27.7 percent pass rate without the external scaffold.

The result is early, but it points to a useful direction for agent training. Instead of treating scaffolds as temporary prompt wrappers, teams may be able to use them as training partners that teach models reusable procedures.