AI ENGVisual Encyclopedia

MODULE 5: REASONING & VERIFIABLE REWARDS · SCENE 14

GRPO: Group-Relative Advantages, No Critic

Sampling a group, normalizing rewards within it, and deleting the value network from the memory bill.

SAMPLE ir = 1.0A = 1.00 ↑ REINFORCECRITIC DELETED · SAVES ~112 GB
GROUP REWARDS (μ = 0.51)1.0+1.00.0-1.01.0+1.01.0+1.00.0-1.00.1-0.81.0+1.00.0-1.0

Advantage is a z-score inside the sampled group: the group mean replaces the learned value function. Compute moves from a critic model to extra rollouts.

TECHNICAL BREAKDOWNModule 5: Reasoning & Verifiable Rewards

GRPO: group-relative advantage, no critic

PPO's value model exists to reduce the variance of a single rollout's return estimate. GRPO deletes it: for each prompt, sample a group of G responses, then use the group's own reward statistics as the baseline. Advantages become a simple normalization inside the group.

Group Baseline

For each prompt, sample G completions, compute rewards, and set A_i = (r_i − mean) / (std + ε). The group mean replaces the learned value function.

What You Save

No critic model in memory, no value-function training instability, no GAE hyperparameters. The memory bill drops by one full model.

What You Pay

G× more rollouts per prompt per step. Compute moves from parameters (critic) to sampling (rollouts) — usually a good trade for reasoning workloads.

MATHEMATICAL FORMULATION · GROUP ADVANTAGE & THE MEMORY BILL
A_i = (r_i − μ_group) / (σ_group + ε); saved_GB = critic_params_B × 16

Deleting a 7B critic saves ~112 GB of training state — often the difference between a feasible and infeasible reasoning-RL run.

REAL-WORLD PRODUCTION ENGINEERING
  • DeepSeekMath introduced GRPO; DeepSeek-R1 scaled it to emergent long reasoning with groups of 8–64 samples.
  • Watch group reward variance: when every sample in a group is right (or wrong), advantages are ~0 and the batch contributes nothing — curriculum on prompt difficulty matters.