What most people call "Benchmaxxing" is not training on test but a different focus on capabilities.
Seen people use this chart of FrontierCode being able to differentiate models better as indicative of benchmaxxing and I think thats not quite accurate.
Some thoughts below :)