Microsoft researchers: LLMs degrade “artifact fidelity”

19 LLMs, 50% error rates on average. Blame the harness?

Updated May 19, 2026 5:50 PM / Published May 18, 2026 5:20 PM 3 min read

Microsoft researchers: LLMs degrade “artifact fidelity” — Image credit: https://unsplash.com/@gaellemarcel

Microsoft researchers warned that “even frontier models” corrupt an average of 25% of document material during extended workflows.

Testing 19 LLMs on a bespoke set of work environments across 52 domains, the researchers found that the models on average degraded 50% of the material they were given to work with, as tasks progressed.

This content is for paying members only

Already have an account? Sign in

Add The Stack on Google

Edward Targett

Ed is a co-founder of The Stack. He previously edited Computer Business Review. He has also covered energy markets. He started his journalism career on local papers. He left school at 15 and has made a living asking "but why?" ever since.