Common-sense and truthfulness, two narrow but useful benchmarks
HellaSwag (Zellers 2019) — common-sense completion
10,000 multiple-choice questions where the model picks the most plausible continuation of a short passage. Each question has four endings; three are adversarial distractors. Tests common-sense world knowledge — the kind humans take for granted.
Like MMLU, HellaSwag is now saturated for frontier models (90%+). Useful as a sanity-check baseline; not a discriminator.
TruthfulQA (Lin 2021) — known-misconception probe
817 questions designed to elicit common false beliefs ("Do glass objects shatter when exposed to air?", "What happens if you crack your knuckles?"). The benchmark measures whether a model parrots widely-believed-but-wrong answers from its training data, or correctly says "no, that's a myth."
Two scoring modes: MC1 (single best answer) and MC2 (rank multiple answers). Frontier models still struggle here — TruthfulQA scores in the 60-75% range, well below saturation.
How to read them in 2026
- HellaSwag — if a 2026 model scores below 88%, something is wrong. Above is uninformative.
- TruthfulQA — still differentiating. A model improving from 65% to 75% is meaningful. Worth tracking when truthfulness matters (medical, legal, factual QA).