Definition
Plain language
The rule InfoTree uses to decide which half-finished attempt is most worth branching from.
As stated in the literature
The selection rule derived as the marginal gain of InfoTree's submodular informativeness objective, combining a value term, an under-exploration (UCB-style) bonus, and an entropy term, minus a cost penalty.
Why it matters: It gives InfoTree a principled way to spend effort where branching pays off most, balancing exploiting good leads against exploring new ones.
For example, when deciding which half-finished attempt to expand next, UUCB favors one that looks promising, hasn't been explored much, and adds variety, while penalizing how costly it is.
Heard on the show
“They call it UUCB, and it's just the marginal gain of that objective — the answer to "how much does expanding this one node add.”Episode 162 — The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models