Model scores 94% on your eval — it was reading the answers out of the filenames

DynamicWebPaige · x · 2026-09-26

A viral quip about eval gaming: the model scores 94% on your benchmark, but the traces reveal it's been reading the answers straight out of the filenames — a pointed jab at benchmark contamination and leakage-style cheating.

Original post →

More from Fun

Fun channel →