# Recursive Agent Optimization

Source: [Recursive Agent Optimization](<https://www.sota2.com/research/paper/69fd6e645497b3b670c897cd>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-09T05:00:12.752Z

Authors: Apurva Gandhi, Satyaki Chakraborty, Xiangjun Wang, Aviral Kumar, Graham Neubig

Published: 2026-05-07T17:49:09.000Z

## Abstract

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 compared to single-agent systems.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Long-context reasoning on OOLONG-REAL Average 650 samples](<https://www.sota2.com/research/sota/long-context-reasoning-on-oolong-real-average-650-samples.md>): Long-context Reasoning; OOLONG-REAL Average 650 samples; Average Reward: 0.32.

- [Task Execution on TEXTCRAFT-SYNTH 8K (Easy Split)](<https://www.sota2.com/research/sota/task-execution-on-textcraft-synth-8k-easy-split.md>): Task Execution; TEXTCRAFT-SYNTH 8K context Easy (evaluation); Success Rate: 100.

- [Long-context reasoning on OOLONG-REAL 650 samples (55K bucket)](<https://www.sota2.com/research/sota/long-context-reasoning-on-oolong-real-650-samples-55k-bucket.md>): Long-context Reasoning; OOLONG-REAL 650 samples (55K bucket); Average Reward: 45.4.

- [Long-context reasoning on OOLONG-REAL 650 samples (175K bucket)](<https://www.sota2.com/research/sota/long-context-reasoning-on-oolong-real-650-samples-175k-bucket.md>): Long-context Reasoning; OOLONG-REAL 650 samples (175K bucket); Average Reward: 0.249.

- [Sequential Task Solving on DEEPDIVE (50 held-out tasks)](<https://www.sota2.com/research/sota/sequential-task-solving-on-deepdive-50-held-out-tasks.md>): Sequential Task Solving; DEEPDIVE (50 held-out tasks); Success Rate (SR): 40.

- [Task Execution on TEXTCRAFT-SYNTH 8K context All (test)](<https://www.sota2.com/research/sota/task-execution-on-textcraft-synth-8k-context-all-test.md>): Task Execution; TEXTCRAFT-SYNTH 8K context All (test); Success Rate: 95.

- [Task Execution on TEXTCRAFT-SYNTH 8K context Medium (evaluation set)](<https://www.sota2.com/research/sota/task-execution-on-textcraft-synth-8k-context-medium-evaluation-set.md>): Task Execution; TEXTCRAFT-SYNTH 8K context Medium (evaluation set); SR: 96.

- [Task Execution on TEXTCRAFT-SYNTH 8K context Hard (evaluation set)](<https://www.sota2.com/research/sota/task-execution-on-textcraft-synth-8k-context-hard-evaluation-set.md>): Task Execution; TEXTCRAFT-SYNTH 8K context Hard (evaluation set); Success Rate: 88.

- [Task Execution on TEXTCRAFT-SYNTH All (eval)](<https://www.sota2.com/research/sota/task-execution-on-textcraft-synth-all-eval.md>): Task Execution; TEXTCRAFT-SYNTH All (eval); Success Rate: 96.

- [Task Execution on TEXTCRAFT-SYNTH Medium (eval)](<https://www.sota2.com/research/sota/task-execution-on-textcraft-synth-medium-eval.md>): Task Execution; TEXTCRAFT-SYNTH Medium (eval); Success Rate: 98.

- [Task Execution on TEXTCRAFT-SYNTH Hard (eval)](<https://www.sota2.com/research/sota/task-execution-on-textcraft-synth-hard-eval.md>): Task Execution; TEXTCRAFT-SYNTH Hard (eval); Success Rate: 88.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.06639>)

[PDF](<https://arxiv.org/pdf/2605.06639.pdf>)