Advanced Reinforcement Learning Interview Questions, issue 19, Feb 14, 2026

The Small-Batch Policy Gradient Trap

Senior RL Engineer interview at OpenAI, and the interviewer asks:

We collected 6 robot trajectories. 5 failed (low reward). 1 succeeded (high reward). We run a vanilla Policy Gradient update on this small batch. What happens to the gradient?

When N is small and rewards are strictly positive, the gradient pushes up both failures and successes, turning learning into a variance-dominated coin flip.

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.