We were amazed by the amount of attention our latest work, “Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win” received at ICML 2025, Vancouver, Canada and want to thank everyone involved in the interesting and inspiring discussions we had.

In short, our paper shows theoretically how succsessful pre-training parameter pruning combined with sparse training regimes (i.e. only updating non-zero weights) is intrinsically connected with expressivity in GNNs. In that context, we also proof SELTH, a novel variant of LTH. Moreover, we provide an in-depth analysis of associated training dynamics and explain certain edge cases where distinguishability of structurally different graphs is irrecoverably lost.

If you wish to know more, check out the conference proceedings, watch our video presentation or feel free to contact us directly.