Definition
Plain language
A setup where a tiny, checkable helper model keeps a bigger AI agent honest by flagging when its story about the world stops matching a grounded prediction.
As stated in the literature
Grounded Iterative Language Planning — pairs an LLM planner with a small trained transition model, gating each step on the overlap between the LLM's predicted state change and the backbone's, and re-prompting the LLM only when they diverge.
Also called: Grounded Iterative Language Planning
Why it matters: It matters because a cheap grounded check can catch an AI planner's wishful mistakes before they derail a whole plan.
For example, if the planner claims an action will unlock a door but the checker predicts it won't, the system stops and asks the planner to rethink.
Heard on the show
“That's GILP — Grounded Iterative Language Planning.”Episode 182 — How a Tiny Model Too Weak to Plan Cuts a Bigger Agent's Hallucinations by 80%