RAO (Recursive Agent Optimization), a paper by Apurva Gandhi, Satyaki Chakraborty, Xiangjun Wang, Aviral Kumar, and Graham Neubig from CMU and Amazon AGI Labs (arXiv:2605.06639), introduces the first reinforcement learning method designed specifically to train recursive LLM agents. Instead of a single agent cramming an entire long-horizon task into one context window, a recursive agent learns when to spawn sub-agents with fresh contexts, how to split subtasks, and how to aggregate their results — effectively learning divide-and-conquer as a trainable policy. Built on an AReAL asynchronous RL backend and a CISPO-style objective, RAO was evaluated on TEXTCRAFT-SYNTH (code synthesis), OOLONG-REAL (long-context reasoning aggregation), and DEEPDIVE (deep research). Recursive agents matched step budgets of 25 per agent versus 100–200 for single-agent baselines, achieving better training efficiency, effective context lengths beyond the native window, generalization to harder unseen tasks, and reduced wall-clock time through parallel sub-agents. The key insight: inference-time scaffolding should not merely wrap the model — the model should be trained to use it.
RAO: Recursive Agent Optimization — Teaching LLM Agents to Delegate Like a CEO
Summary
RAO (Recursive Agent Optimization), a paper by Apurva Gandhi, Satyaki Chakraborty, Xiangjun Wang, Aviral Kumar, and Graham Neubig from CMU and Amazon AGI Labs (arXiv:2605.06639), introduces the first reinforcement learning method designed specifically to train recursive LLM agents. Instead of a single agent cramming an entire long-horizon task into one context window, a recursive agent learns when to spawn sub-agents with fresh contexts, how to split subtasks, and how to aggregate their results — effectively learning divide-and-conquer as a trainable policy. Built on an AReAL asynchronous RL backend and a CISPO-style objective, RAO was evaluated on TEXTCRAFT-SYNTH (code synthesis), OOLONG-REAL (long-context reasoning aggregation), and DEEPDIVE (deep research). Recursive agents matched step budgets of 25 per agent versus 100–200 for single-agent baselines, achieving better training efficiency, effective context lengths beyond the native window, generalization to harder unseen tasks, and reduced wall-clock time through parallel sub-agents. The key insight: inference-time scaffolding should not merely wrap the model — the model should be trained to use it.
This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177619776