Glossary · Term

G2PO

← all terms

Definition

Plain language

A training method that treats an AI agent's many attempts as one shared map instead of separate runs, squeezing more learning out of the same data.

As stated in the literature

Group-Graph Policy Optimization — a GRPO extension that fuses repeated states across rollouts into a graph, pooling per-node value estimates and scoring each action by its global edge-level advantage (temporal-difference gain).

Also called: Group-Graph Policy Optimization

Why it matters: By sharing information across overlapping situations, it extracts more reliable learning signal from the same amount of agent experience.

For example, if an agent reaches the same screen in five different attempts, G2PO pools what happened next across all five instead of judging each attempt alone.

Heard on the show

“And that grading-on-a-curve idea is exactly what G2PO inherits and then pushes further.”
Episode 165 — A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants

Mentioned in 1 episode

  1. 165
    A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants

Related terms