Just like the curve to points node, this mode outputs point locations based on the input curve's evaluated points (the points you see in the viewport). This is expected to be faster, since it doesn't have to sample equal positions on the result, and it's also consistent with the existing choices in the curve to points node.