Researchers develop 'Red Queen Gödel Machine' for co-evolving AI agents and evaluators

Addressing the Evaluation Ceiling in AI

Researchers have developed a new method to overcome a significant challenge in the development of self-improving AI agents. Currently, AI agents that improve themselves by editing and testing their own code often reach an "evaluation ceiling." This occurs because their improvement is limited by fixed evaluators, benchmarks, or test suites, causing progress to slow or stop once the agent has learned everything the fixed signal can distinguish.

Co-Evolving Agents and Evaluators

The new method, termed the 'Red Queen Gödel Machine,' involves both the self-improving AI agent and its evaluator evolving concurrently. Instead of an agent improving against a static test, the evaluation system adapts alongside the agent. As the agent's capabilities advance, the evaluation criteria become more challenging, ensuring a continuous cycle of improvement.

Mechanism of the Red Queen Gödel Machine

The Red Queen Gödel Machine searches through multiple versions of an AI agent while simultaneously improving the evaluator. During each phase, the evaluator remains fixed to allow for reliable progress measurement. At specific checkpoints, a stronger evaluator can replace the older one if it performs better on trusted ground-truth examples. Scores from the previous evaluator are then removed, guiding the next phase with the new, more demanding standard, thus creating a self-improving loop.

Potential Impact and Cost Reduction

This co-evolutionary approach not only aims to enable continuous self-improvement for AI agents but also suggests a way to reduce the computational resources required for their development. The research team, which includes collaborators from NVIDIA and Flower Labs, has published the technical details and initial results in a pre-print paper titled "The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators" on arXiv.