What LLM Benchmarks Actually Say About Prompt Format?

Prompt-format studies and structured-output benchmarks measure different capabilities. This paper distinguishes format sensitivity, Markdown readability, schema compliance, and structured extraction across evaluations including MDEval, JSONSchemaBench, and LLMStructBench.

It examines what those results can support about Markdown, JSON, XML, YAML, and plain text, then proposes workload-specific evaluation. Speed and quality are treated as separate dimensions rather than a single winning format.