Comment by bogtog
6 months ago
A lot of the comments express some type of deliberate cheating the benchmark. However, even without intentionally trying to game it, if anybody can repeatedly take the same test, then they'll be nudged to overfit/p-hack.
For instance, suppose they conduct an experiment and find that changing some hyper-parameter yields a 2% boost. That could just be noise, it could be a genuine small improvement, or it may be a mix of a genuine boost along with some fortunate noise. An effect may be small enough that researchers would need to rely on their gut to interpret it. Researchers may jump on noise while believing they have discovered true optimizations. Enough of these types of nudges, and some serious benchmark gains can materialize.
(Hopefully my comment isn't entirely misguided, I don't know how they actually do testing or how often they probe their test set)
I cringe every time I see "my IQ increased by X points after doing Y" posts on Twitter - yes, you had a practice run on Raven's progressive matrices a month ago, that helped, these have a limited question bank and the effect of Y is marginal. That said, obviously, test taking is a skill (separate from background knowledge and both general/domain-specific ability) and should be trained if you expect to have life-altering events based on tests (i.e., do an LSAT course if you want to go to law school). Conversely, shouldn't be done if you think it will limit you through superstition ("I had a score of X, thus I can only perform around level of X+fudge factor"). For an LLM company a good test score is a valuation-altering event!