MMLU (Massive Multitask Language Understanding) — tests a different competence than → SuperGLUE

Both function in this corpus as regression-catching suites -- benchmarks a paper runs, unmodified, to check that its intervention did not quietly break something else -- but they probe different competence. SuperGLUE's RTE and WSC tasks, the two InstructGPT reports results on, test whether a model can perform specific linguistic inference: recognizing textual entailment and resolving pronoun reference (instructgpt, §"D Automatic evaluation details", p. 44). MMLU tests something closer to stored knowledge and subject-matter reasoning across 57 independent domains, from moral scenarios to professional law, scored by the probability assigned to the correct multiple-choice answer (contrastive-activation-addition, §"7 Effect of CAA on general capabilities", p. 6). A method could hold one steady while damaging the other -- steering a behavior vector into the residual stream is exactly the kind of intervention where a narrow linguistic skill and broad factual recall might fail independently, which is why a capability check plausibly needs both, not just whichever benchmark a paper happens to already be running.