benchmaxx

From benchmark +‎ -maxx.

• IPA^(key): /ˈbɛnt͡ʃmæks/

benchmaxx (third-person singular simple present benchmaxxes, present participle benchmaxxing, simple past and past participle benchmaxxed)

1. (artificial intelligence, informal, derogatory) To optimize a AI model's performance on a specific benchmark instead of general capability.
   • 2025 December 15, Ben Dickson, “Why AI benchmarks are broken”, in TechTalks‎^[1], archived from the original on 7 March 2026:

     There are different ways companies are benchmaxxing. Some of them are outright cheating, like directly training on benchmarks.

   • 2025 December 19, Andrej Karpathy, “2025 LLM Year in Review”, in karpathy.bearblog.dev‎^[2], archived from the original on 21 February 2026:

     In the typical benchmaxxing process, teams in LLM labs inevitably construct environments adjacent to little pockets of the embedding space occupied by benchmarks and grow jaggies to cover them. Training on the test set is a new art form.