Researchers at Cambridge's Machine Learning Systems Lab built a framework that lets AI agents keep improving themselves without stalling against a fixed test, the university said. Self-improving agents typically plateau once they've learned everything a static evaluator or test suite can teach them, according to the announcement.
The team's fix, called the Red Queen Gödel Machine, evolves the evaluator at the same time as the agent it grades. "Instead of improving an agent against a fixed test, we let the evaluation evolve alongside the agent," said Alex Iacob, the PhD student who led the work under Professor Nic Lane, with collaborators from NVIDIA, Flower Labs, MBZUAI and Inria, the university said.
In early results on scientific paper writing and mathematical proof grading, agents that co-evolved with their evaluator reached acceptance rates 1.78 to 1.86 times higher than previous self-improving approaches, according to Cambridge. Mixing cheaper open source models in with expensive frontier models cut computational costs roughly 13-fold while holding performance steady, the university said.
The name references the Red Queen from Lewis Carroll's "Through the Looking-Glass" and the evolutionary biology hypothesis that species must keep adapting just to stay in place against rivals that are adapting too, according to Cambridge.
For teams building agents that grade or improve their own outputs, a fixed rubric is the plateau this research is aimed at. A test that evolves alongside the thing it's testing is a different design problem than the static eval suites most teams reach for first.