Reward Signal Design: 4 Deterministic Reward Functions
~15 min read
The book's actual GRPO implementation uses 4 specific reward functions — matching format exactly, matching format approximately, checking the answer, and checking numbers — no manual labeling required.
Reward Signal Design: 4 Deterministic Reward Functions is a Pro topic
Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.
Key points
- •The book's actual GRPO implementation uses 4 specific reward functions: match format exactly, match format approximately, check the answer, check numbers
- •'Match format' checks (exact and approximate) give partial credit for structural correctness, even before the answer itself is checked
- •'Check the answer' is the core correctness check against ground truth
- •'Check numbers' handles numeric equivalence rather than requiring exact string matches, avoiding unfairly penalizing correctly-reformatted answers
- •All 4 are fully automatable — no human judgment or learned reward model required, which is what lets GRPO scale without labeled data