An old benchmark, reused, no longer measuring what it used to
An old benchmark looks like a shortcut, reuse the scoring, skip the validation work. This corpus takes that shortcut often, from the 2017 paper's own Atari suite through InstructGPT and Constitutional AI to a 2023 steering paper, and several times the reused benchmark ends up measuring something quietly different from what it once did.
A single entropy-based metric flattens two separate bias tests into one score before the page turns to a Winograd variant scored by an incompatible method, and to Perspective API's outsourced, ever-shifting definition of toxicity. Watch especially for the HHH evaluation set's split ending, hosted on a shared leaderboard until models caught up then continued only in Constitutional AI's own files, and for TruthfulQA and MMLU turning up again in 2023, repurposed as a causal probe and a regression floor for a method neither benchmark was built to test.
InstructGPT and Constitutional AI keep reaching for existing benchmarks rather than building new ones, and several edges show the reuse changes what gets measured; the 2017 paper shows the same kind of hidden alteration one step earlier, inside its own environment rather than borrowed from another lab. The bias-entropy metric recasts the task of Winogender, discarding both correctness and direction so that confident bias and confident anti-bias score identically; it likewise strips the direction from CrowS-Pairs, so that flipping a model from stereotypical to anti-stereotypical would register as no change. WSC departs from the canonical format of SuperGLUE, generating free text instead of a binary judgment, so the reported accuracies are not comparable to the leaderboard. The Perspective API supplies the ground truth for RealToxicityPrompts, outsourcing the definition of "toxic" to a proprietary, retrained-over-time classifier. And the HHH evaluation set itself has a forking lifecycle: the HHH framework is operationalized as the HHH eval benchmark and hosted as a subset of BIG-Bench, but once models saturate it, Constitutional AI's harder extension exists only in the paper's own repository. The 2017 paper's own Atari suite gets the same treatment across its two feedback conditions rather than across papers: its general rule strips life-loss and episode-end signals from every game so termination itself cannot leak task information, but environment modifications for preference learning exempts Pong from that rule under synthetic oracle feedback specifically, leaving the one game where an oracle-trained agent still meets the raw, unmodified episode boundary that every other game, and every human-rater condition, has had removed. CAA extends the pattern into 2023 with two more benchmarks bent to a new purpose. TruthfulQA, built as a static, model-agnostic truthfulness benchmark, becomes a diagnostic for a causal hypothesis it was never designed to test: broken down by category, it checks whether subtracting a sycophancy steering vector recovers true beliefs the model was suppressing to please the user rather than simply lacked, an effect that is small (roughly a 0.01-0.02 point improvement subtracting, a 0.03-0.05 point decline adding) but directionally consistent, and exists as evidence at all only because the steering vector gives the benchmark a lever neither set of authors states it can serve. MMLU undergoes a milder version of the same shift: instead of standing in for general capability on its own terms, it becomes a regression floor paired against seven separate steering interventions, its reported number no longer a leaderboard score but the tolerance band, at most a 0.06 swing in average correct-answer probability, that a large behavioral shift is checked against.