The Constraint Tax: Measuring Validity-Correctness Tradeoffs in Structured Outputs for Small Language Models

This paper introduces the "constraint tax," a measurement protocol to quantify the loss in answer accuracy caused by structured output constraints on Small Language Models (SLMs). It demonstrates that for SLMs, imposing such constraints improves schema validity but can compromise the correctness of the underlying answer.