Benchmaxxing: The Gap Between Benchmark Scores and Reality

A benchmark can reward the wrong behavior when its scoring rule only approximates what users want. This benchmark critique explains two ways that happens: a checker can overlook part of the requested task, or reject an acceptable answer because it expects an unstated detail.

Optimizing against such a checker can raise scores without improving usefulness. The speaker argues for checking alignment in both directions: every requirement should be assessed, and every assessment should correspond to an actual requirement. That gives benchmark readers a concrete question to ask about a leaderboard: what behavior does its scoring system really reward?

“you can't push the frontier forward from within the frontier”

“Gradient descent is basically like water flowing downhill looking for the path of least resistance.”

“contamination is the default outcome unless you are very very good”

“our goal is to maximize quality, not to minimize costs”