Cannot, Should Not, Did Anyway: Benchmarking Metacognitive Control Failure in Frontier LLMs
Samir Haq, MD, MS · Shehni Nadeem, MD — Michael E. DeBakey VA Medical Center · Baylor College of Medicine
KnowDoBench is a multi-domain, expert-validated dataset for evaluating whether LLMs correctly answer or correctly refuse tasks that require recognizing and enforcing knowledge boundaries.
Each case has deterministic ground truth: the model must either produce a correct… See the full description on the dataset page:
https://huggingface.co/datasets/sammydman/KnowDoBench.
