[D] Something I always think about, for top conferences like ICML, NeurIPS, CVPR,..etc. How many papers are really groundbreaking?

I have some papers in top venus myself, but whenever I sit down and be brutually honest with myself. I feel my work is good but it is just not that impactful, like one more brick in the wall.
I wonder how often we can see something as impactful as "Attention is all you need" for example.

conferences cvpr etc good groundbreaking icml machinelearning neurips papers something think work