RAO paper says recursive training on easy tasks can generalize to harder ones

a1zhang · x · 2026-07-27

This repost discusses the paper Recursive Agent Optimization (RAO) and a related finding about how agents can generalize better.

The core claim is that training only on shorter, easier tasks with a recursive / divide-and-conquer structure can help models generalize to much longer and harder tasks.
The authors also observed that training is more efficient when the setup exploits the recursive structure of inference.
In particular, it helps to train and reward sub-agents, not just the root agent.
The post also connects RAO with other ideas in the same direction, including Reasoning Cache and Context Folding, emphasizing better use of inference-time compute for extrapolation across domains.

More from coding & agent