How it works
ZK-rollups bundle transactions into batches and execute them off chain. Instead of sending every transaction to layer 1, ethereum.org explains, operators submit a summary of the changes needed to represent all the transactions in the batch. Alongside it goes a validity proof, a cryptographic assurance that the proposed state change really is the result of executing that batch.
What the proof does
The proof is checked by a contract on Ethereum. If it verifies, the new state is accepted. If it does not, the batch is rejected. Verification is cheap compared with re-running the batch, which is what makes the approach scale. The proof reveals nothing extra about the transactions; 'zero-knowledge' refers to that property, though scaling is the main reason rollups use it.
Faster exits
Because correctness is proven before acceptance, there is nothing to dispute afterwards. ethereum.org states there are no delays when moving funds from a ZK-rollup to Ethereum, since exit transactions are executed once the contract verifies the validity proof. By contrast, withdrawing from optimistic rollups is subject to a delay to allow challenges.
Trade-offs
Producing proofs takes heavy computation, which adds cost and complexity for operators. Making proofs for general-purpose EVM code is hard, so designs vary in how closely they match Ethereum's own behaviour and how mature they are. When you use one, the practical differences are faster finality to Ethereum and sometimes slight differences in how applications behave. As with any layer 2, read the documentation on who operates the system and what happens if they stop.