Definition
Plain language
A way of collecting an AI's practice attempts as a branching tree, spending effort where the reasoning actually forks instead of on independent tries.
As stated in the literature
A rollout-collection method that grows a shared tree of attempts and selects nodes to expand by the marginal gain of a submodular informativeness objective (UUCB), provably near-optimal via the greedy guarantee.
Why it matters: It concentrates effort on the moments that actually decide outcomes, getting more useful learning out of the same compute budget.
For example, instead of generating ten separate solutions from scratch, it grows one tree of attempts and adds new branches only where the reasoning genuinely splits.
Heard on the show
“They run their system — they call it InfoTree — against that same axis.”Episode 162 — The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models