The Rejection Sampling Trap

Senior ML Engineer interview at Google DeepMind, and the interviewer asks:

“Your lead wants to skip RL entirely: sample from the model, keep the rollouts that pass the tests, fine-tune on them, repeat. When is that good enough to ship, and where does it hit a ceiling that policy gradient doesn’t?”

Don’t say: “Filtering by reward is basically RL, so it’ll work the same.”