DeepSeek V3.2 schema checks fail subtle factual errors

An author tests how well constrained decoding and JSON schemas prevent language models from generating incorrect data. By running 1,600 synthetic documents through DeepSeek V3.2, the author finds that while rules successfully catch formatting errors and structural mistakes, they often fail to prevent well-formed answers that contain subtle factual errors. Open article →