Optimizing a model or prompt mainly to raise benchmark scores.
Benchmaxxing is like memorizing the answer key before a school test. Great grade, then it messes up a pizza order.
It can make leaderboard scores look better. It can also make model shopping harder.
Benchmark contamination
Benchmark contamination makes benchmaxxing easier and harder to spot.
Leaderboard
Benchmaxxing usually focuses on climbing the leaderboard.
MMLU
A famous benchmark like MMLU is more likely to get special tuning.
Reward Hacking
Benchmaxxing is reward hacking for tests. It chases points, not real skill.