Patrick Baitman retweetledi
Most systems that let an AI agent split a hard problem into pieces and hand those pieces to fresh copies of itself treat that splitting as scaffolding bolted onto a model that was never trained to do it, so the model is left guessing about when delegating helps, how to phrase a subtask, and how to combine what comes back.
This 2026 paper from Amazon and CMU trains a single language model to make those decisions itself: the same model runs at every node of a tree of delegated subtasks, and learning happens through reinforcement learning.
The central design choice is a local reward—each node is scored on whether its own assigned task succeeded plus a bonus for how often its delegated children succeeded, which lets the model learn decomposition even before it can solve the full task.
An agent trained with an 8K-token context window solves crafting tasks that exceed that window (95% versus 24% for a non-recursive baseline), a 30B model approaches the long-document performance of larger frontier systems, and the recursion the model learns is adaptive, going deeper on harder problems and staying shallow when one round of splitting suffices.
Read with an AI tutor: chapterpal.com/s/2b7c9ae1/rec…
PDF: arxiv.org/pdf/2605.06639
English
