GPAI.WIKI
open problems/OP-003··contested

A measure of generality that benchmarks cannot be trained on

Status
contested
Difficulty
conceptual
Opened
2026-08-22

Why it is contested rather than merely open

Some hold that this is unsolvable in principle — that Goodhart’s law applies to any published measure, and that the correct response is a rotating private evaluation regime rather than a better metric. Others hold that a measure defined over task-acquisition cost rather than task performance would be robust, because optimising against it is indistinguishable from genuinely improving.

The archive files this as contested because both positions have serious proponents and the disagreement is substantive, not terminological.

What would count as progress

  • A measure with a stated invariance property, plus evidence it survives direct optimisation.
  • A formal argument that no such measure exists, which would settle it the other way.
work on this

Partial results, negative results and sharper restatements are all welcome. Open a pull request that edits this file, or add a paper or implementation entry that references OP-003.

reactions
no reactions yet
react on GitHub