You're right, I did some reading: “When a new block is found, mining pools send miners a block template with no transactions so they can start searching for the next block as soon as possible. They send a block template full of transactions right afterward, but a full block template is a bigger data transfer and takes slightly longer to reach miners.”

Basically, since it takes a few seconds longer to send a full block template, miners hash away at an empty template and occasionally find a new empty block...

Once the fee's overtake the coinbase value by an order of magnitude, this practice will stop since it makes more business sense to always hash a more valuable template.