A training loop calls loss.backward() and optimizer.step() but never optimizer.zero_grad(). What happens?
Answer: B. Gradients accumulate across steps, so updates grow and training becomes unstable
.grad is summed into on every backward pass. Without zeroing, each step uses the sum of all previous gradients.