TeXFix-Bench provides a multi-format benchmark for LaTeX, Typst, and Markdown repair based on a 18-category taxonomy of real-world faults. Experiments show that DocMut faults, grounded in mined human errors, are 5.6-9.2 percentage points harder to repair than standard synthetic mutations.
HOW THIS AFFECTS YOU
●
builderThis offers a more realistic testing ground for LLMs tasked with technical document automation.
●
researcherThe grounded-theory approach provides a more robust fault model than ad-hoc synthetic edits.