Researchers from NVIDIA have introduced a new method for recursive self‑improving AI agents that can continue to enhance their own code without hitting a fixed evaluation ceiling.

Team member Alex Iacob, a PhD student in the Machine Learning Systems Lab under Prof Nic Lane, explains that a self‑improving agent is limited by the test that scores it, which defines its progress and creates a ceiling.

The Red Queen Gödel Machine co‑evolves the agent and its evaluator, letting the test become harder as the agent improves, creating a rising bar that keeps performance moving forward. Results show co‑evolved paper writers achieve 1.86× higher acceptance rates and co‑evolved graders reach 9% higher ground‑truth accuracy, while a hybrid setup using NVIDIA Nemotron 3 Ultra and ChatGPT cuts search‑token costs by about 13×, demonstrating a path to more capable open agents at lower cost.

Although the work is still preliminary, the team notes that longer search horizons are required to gauge scalability, and they plan to open‑source the method to encourage wider adoption while testing it on a broader range of AI tasks.