MISP-Bench decomposes LLM misinformation damage under user-provided false priors.
2,494 multiple-choice items (2,194 MedMCQA + 300 GSM8K) under 14 prompt
conditions, designed to isolate which structural component of a wrong user
prior — the answer, the rationale, or their combination — drives downstream
model error, and to test whether common safety prompts ("verify the reasoning
first") actually mitigate it.
The audited corpus (1,724 items) is materialized at evaluation… See the full description on the dataset page:
https://huggingface.co/datasets/yh0502/misp-bench.
