Benchmaxing is defined as targeted post-training attempts to optimise for higher benchmarks. This doesn’t change the aggregate intelligence or capability of the model, so it can make a model appear more capable than it is.