Definition
Plain language
An earlier method for training AI agents that noticed repeated situations but still judged each one from a single run.
As stated in the literature
A group-relative agent-RL method that detects recurring states but estimates each from one trajectory's outcome and compares actions only locally; the baseline G2PO improves on with graph-level pooling and global advantage.
Why it matters: It marks the step before pooling repeated experiences, clarifying why combining information across runs gives a more stable training signal.
For example, it notices an agent revisited the same menu in several runs but still rates that menu's choices using only what happened in one run.
Heard on the show
“… Even the closest prior method — GiGPO — noticed that identical states recur, but it still judged each state from one trajectory's outcome, …”Episode 165 — A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants