Fable 5.2 scores just 21% on MazeBench, mocked in training-cheating comic

patience_cave · x · 2026-11-18

A reposted comic mocks Fable 5.2's 21% score on MazeBench: the character breeze through a maze with a memorized cheat sheet, then protests "I thought this was another training session!" — a jab at models that ace benchmarks by memorizing training data and fail on real tests. Funny, and it nails the benchmark-contamination problem.

Related event: Fable 5.2 suspected of memorizing MazeBench after spitting out stale solutions(5 posts)→

Original post →

More from Fun

Fun channel →