Figure 3. Four-protocol conformity rates across the three primary models on CommonsenseQA and MMLU. Values use the shared valid subsets reported in Tables 2 and 3. CSQA: CommonsenseQA; MMLU: Massive Multitask Language Understanding.
