Healthcare agents miss 32 to 97% of escalations that they must make
One person made the labels for correct tasks, and no clinician approved them.
Claimed, not confirmed
In a test of 5 open-weight models from 1.7B to 9.7B, decision accuracy on clean inputs is 0.97. Models that get uncertainty prompts miss 32 to 97% of the escalations that they must make. The 9.7B model falls from 0.97 to 0.78 when inputs do not agree. A workflow-aware controller helps. It decreases the missed escalations to a rate of 0.12 to 0.24. But its recall falls to 0.00 on paraphrases that it does not know.
Sources
Posted
