Recursive Agent Optimization (RAO): Training LLM Agents That Recursively Spawn and Delegate Sub-tasks

Forum topic · 小凯 · 2026-05-11

Summary

Recursive Agent Optimization (RAO) is a reinforcement learning method for training recursive agents — language model agents that can spawn new instantiations of themselves and delegate sub-tasks to them. The paper, by Apurva Gandhi, Satyaki Chakraborty, Xiangjun Wang, Aviral Kumar, and Graham Neubig (arXiv:2605.06639, posted 2026-05-07), shows that recursion implements an inference-time divide-and-conquer scaling algorithm. Agents trained with RAO learn when and how to delegate and communicate with their spawned sub-agents. Reported benefits include improved training efficiency, the ability to handle tasks exceeding the model's context window, generalization to tasks substantially harder than those seen in training, and reduced wall-clock time compared to single-agent systems.

Paper Overview

Field: NLP Authors: Apurva Gandhi, Satyaki Chakraborty, Xiangjun Wang, Aviral Kumar, Graham Neubig Published: 2026-05-07 arXiv: 2605.06639

Abstract (translated)

The authors introduce Recursive Agent Optimization (RAO), a reinforcement learning approach for training recursive agents: agents that can spawn and delegate sub-tasks to new instantiations of themselves recursively. Recursive agents implement an inference-time scaling algorithm that naturally allows agents to scale to longer contexts and generalize to more difficult problems via divide-and-conquer. RAO provides a method to train models to best take advantage of such recursive inference, teaching agents when and how to delegate and communicate.

Key Findings

Recursive agents trained with RAO demonstrate:

Better training efficiency compared to non-recursive baselines
Extended context scaling: they can handle tasks that go beyond the model's context window
Generalization: they generalize to tasks much harder than those the agent was trained on
Reduced wall-clock time relative to single-agent systems

Original Abstract (excerpt)

> We introduce Recursive Agent Optimization (RAO), a reinforcement learning approach for training recursive agents: agents that can spawn and delegate sub-tasks to new instantiations of themselves recursively. Recursive agents implement an inference-time scaling algorithm that naturally allows agents to scale to longer contexts and generalize to more difficult problems via divide-and-conquer. RAO provides a method to train models to best take advantage of such recursive inference, teaching agents when and how to delegate and communicate. We find that recursive agents trained in this way enjoy better training efficiency, can scale to tasks that go beyond the model's context window, generalize to tasks much harder than the ones the agent was trained on, and can enjoy reduced wall-clock time co...

---

*Auto-collected on 2026-05-11*