Large benchmark sets like GMTKN55 [1] let us analyse the performance of density functional theory over a diverse range of systems and bonding types. However, routine calculation of over 1500 relative energies and 2500 single point energies is (a) a costly and (b) can potentially obscure large errors through averaging, giving a false sense of security. Previous work developed the 'Diet’ benchmark sets consisting of 30, 100, and 150 relative energies, which is capable of reproducing the key results of the full GMTKN55 database, addressing point (a). [2] This work introduces the `Poison’ benchmark sets, consisting of 30 relative energies and representing some of the most difficult-to-model systems in the GMTKN55 database, addressing point (b). [3] Discussion will focus on the results of these poison benchmark sets, which show that general GMTKN55 rankings are not necessarily consistent with poisoned rankings. These suggest that common benchmarking strategies may need to be revisited when modelling difficult chemical systems.

1. Lars Goerigk, et al. Phys. Chem. Chem. Phys. 2017, 19, 32184
2. Tim Gould. Phys. Chem. Chem. Phys. 2018, 20, 27735
3. Tim Gould and Stephen G. Dale. Phys. Chem. Chem. Phys. 2022 (Accepted)