Judgment models are a class of models that answer semantic questions by scoring a finite set of candidate outcomes. They work well for local decisions, but a single scoring step may not be enough for questions that require several intermediate inferences.
The ABACUS framework, introduced in a new arXiv preprint, explores whether reusable programs can turn these local decisions into powerful multistep reasoning. The core idea is to guide the reasoning process with programs, while keeping the underlying judgment models frozen — that is, not updating their parameters.
The abstract describes ABACUS as a framework that jointly constructs something, but the sentence is cut off in the available text. As a result, the exact mechanism of that joint construction is not specified here. What is clear is the ambition: to compose simple judgments into longer chains of reasoning without retraining the models.