DeepInfer xgrammar 0.2.2 miscompiles JSON Schema in structured output

The author examines how structured-output decoders like DeepInfer handle JSON Schema, finding that certain keywords are enforced while others are compiled and let through or rejected incorrectly. To bridge this gap, the author proposes a bundle format that combines input rules, shapes, and output rules into a single file to ensure reliable LLM inference. Open article →