Skip to lesson
Exit
Deep Dive: Graders, Rewards & Evals1 / 2

1 min lesson

Cost versus fidelity is a live tradeoff

Use "Cost versus fidelity is a live tradeoff" to explain each part and the role it plays.

Step 1 of 2

Cost versus fidelity is a live tradeoff

Every rollout calls the grader, so the grader's cost multiplies across the whole training run. A large-model judge that scores beautifully may be too slow to run on every sample, starving the policy of gradient. A cheap heuristic runs fast but rewards the wrong thing.

The dials you trade against each other
Grader latency / cost
Sets rollout throughput. A 5-second judge on millions of samples can dominate the compute bill.
Grader fidelity
How well it tracks true quality. Higher fidelity usually means a bigger, slower model.
Gate-then-judge
Run the cheap execution check first; only pay for the expensive judge on samples that survive.
Reward-model distillation
Train a small RM on the big judge's labels, then serve the cheap RM in the loop.

Naming this tradeoff out loud - and the gate/distill tricks that ease it - reads as someone who has run the pipeline, not just designed one.

Interview move

Asked to design a grader for, say, "refactor this module," don't hand-wave "an LLM judges it." Walk the stack: an execution gate to kill broken builds, a rubric-driven judge for readability and minimal diff, a human-labeled eval set to calibrate the judge and a cost plan (distill the judge into a small RM) so it survives the rollout budget. Then name how you'd attack your own grader.