Study finds LLMs corrupt documents during delegated editing workflows, with frontier models averaging 25% content degradation

Summary

This paper introduces DELEGATE-52, a benchmark to evaluate how well Large Language Models (LLMs) handle delegated document editing tasks across 52 professional domains. Testing 19 LLMs, the study finds that current models systematically degrade documents during long workflows, with even top-tier frontier models corrupting an average of 25% of document content by the end of extended interactions. The research reveals that agentic tool use does not improve performance, and degradation worsens with larger documents, longer interactions, and distractor files. The authors conclude that current LLMs are unreliable delegates that silently introduce sparse but severe errors that compound over time.