A sealed final test doesn't stop agents from overfitting the validation loop

rhythmisbackUwU · reddit · 2026-08-19

Discussing the AQuA paper's evaluation design, the author notes the sandbox uses fixed splits and an evaluator, with the system receiving validation scores during search and adapting to that feedback; the final test runs only once after configuration is frozen. He argues this is a procedural rather than cryptographic boundary—it protects the reported final number but leaves the search process exposed to adaptive overfitting of the visible validation slice. Final-test isolation and search validity are separate properties.

He asks: what additional evaluation layer would best detect when a long agent search has started optimizing the validation slice rather than the underlying task?

Original post →

More from coding & agent

coding & agent channel →