Tag
The paper proposes ModelLog, a declarative probabilistic framework for evaluating language models by defining semantic constraints over token predictions, linking evaluation to learning through shared semantics.