38% of correct LLM markdown has broken fences

Correct-looking markdown from large language models breaks its own code fences, a new paper finds. Tools that regex-split on fence markers, or feed raw model output straight into a renderer, fail silently when that happens. The model's answer can be entirely correct while the code extraction step still breaks.

LatentMD ran 9 language models through about 37,600 generations, scoring content correctness separately from fence-boundary correctness. Even when the content was right, the boundaries were broken 38% of the time, per LatentMD's paper.

The failures aren't about nesting depth. They cluster around same-family symmetric delimiters colliding, such as a triple-backtick fence appearing inside content that already contains backticks. The same pattern shows up in Python's triple-quote docstrings, another symmetric delimiter, while JSON's asymmetric brackets and braces stayed steady across those tests.

The paper doesn't say which of the 9 models fared best, so there's no way to just pick a safer one.