New Preprint Finds LLMs Fail to Reliably Use Uncertainty Markers for Confidence