Red Queen Gödel Machine co-evolves AI agents and their evaluators

A team at Cambridge's Machine Learning Systems Lab, with collaborators from NVIDIA, Flower Labs (a departmental spin-out), MBZUAI and Inria, has published a pre-print method called the Red Queen Gödel Machine. It targets a known limit in recursive self-improving AI agents: agents that edit their own code, test variants and keep what scores better usually stall once they have learned everything a fixed evaluator, benchmark or test suite can distinguish. First author Alex Iacob, a PhD student supervised by Prof Nic Lane, put it directly: "A self-improving agent can only get as good as the test that scores it. The test does not merely measure progress, it defines it, so the efficacy of the test becomes a ceiling the agent cannot climb past."

The Red Queen Gödel Machine addresses this by having the agent and its evaluator evolve together rather than testing the agent against a static signal. "Instead of improving an agent against a fixed test, we let the evaluation evolve alongside the agent," Iacob said. "As the agent gets better, the evaluation also gets harder, and the bar keeps rising." In practice, the system searches through many candidate versions of an agent while separately improving the evaluator that judges them. Within each phase the evaluator stays fixed so progress can be measured reliably; at checkpoints, a stronger evaluator can replace the old one if it performs better on trusted ground-truth examples, and scores produced by the old evaluator are discarded so the next phase is judged by the tougher new standard. The name references the Red Queen from Lewis Carroll's Through the Looking-Glass, who tells Alice "it takes all the running you can do, to keep in the same place," and the Red Queen hypothesis from 1973 evolutionary biology, which holds that species must keep adapting just to survive against other species that are also evolving.

The team tested the framework on scientific paper writing and reviewing, and on Math Olympiad-level proof writing and grading. In the paper's own words, "co-evolved scientific paper writers reach 1.78x-1.86x higher acceptance rates under a diverse agent-as-a-judge panel, while co-evolved graders reach 9% higher ground-truth accuracy." Both figures are relative gains over the paper's baseline; the source does not give the underlying absolute acceptance-rate or accuracy numbers.

A separate finding concerns cost. In one experiment, the researchers co-evolved AI reviewers and writers of scientific papers using NVIDIA's open Nemotron 3 Ultra model alongside ChatGPT-5.5. In the paper-reviewing task, this hybrid setup approached, though did not match or exceed, the performance of ChatGPT-5.5 used on its own, while cutting search-token costs by about 13 times. Co-author Prof Nic Lane cautioned against reading too much into this: "This is a narrow result, and we should be careful not to overstate it. But it also indicates where this method can go. If open models can carry the bulk of the search while stronger frontier models guide the higher-level improvement process, then we have a plausible path toward much more capable open agent systems at far lower cost." Daniel Burkhardt, NVIDIA's Developer Relations Manager and a co-author on the paper, added: "This effort shows how open models can play an important role in advanced agentic systems. NVIDIA Nemotron models are designed to be efficient and capable for reasoning and agent workloads, and this work shows how they can be used as part of a broader self-improvement loop."

The researchers describe the work as preliminary and say longer search horizons will be needed to understand how far the approach can scale. The method has not been released yet; the team says it will be open sourced to support wider adoption, and plans further experiments across a broader range of AI tasks at larger scale. The paper, titled The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators, is on arXiv, with Iacob and Lane among a 13-author list that also includes Burkhardt, Andrej Jovanović, William F. Shen, Meghdad Kurmanji, Nurbek Tastan, Lorenzo Sani, Niccolò Alberto Elia Venanzi, Ambroise Odonnat, Zeyu Cao, Bill Marino and Xinchi Qiu.

Key facts

The Red Queen Gödel Machine co-evolves a self-improving AI agent with its own evaluator, so the evaluation gets harder as the agent improves instead of capping progress at a fixed benchmark.
Co-evolved scientific paper writers reached 1.78x-1.86x higher acceptance rates under a diverse agent-as-a-judge panel; co-evolved graders reached 9% higher ground-truth accuracy on Math Olympiad-level proof grading.
Pairing NVIDIA's open Nemotron 3 Ultra with ChatGPT-5.5 in the paper-reviewing task approached ChatGPT-5.5's solo performance while cutting search-token costs by about 13 times.
First author Alex Iacob is a PhD student in Cambridge's Machine Learning Systems Lab under Prof Nic Lane; co-authors include NVIDIA's Daniel Burkhardt, with additional support from Flower Labs, MBZUAI and Inria.
The method is a pre-print on arXiv, not yet released; the team plans to open source it and calls the hybrid-cost result preliminary, pending tests over longer search horizons.

Why it matters

Self-improving AI agents that edit and test their own code have generally been bounded by whatever fixed evaluator, benchmark or test suite scores them: once the agent learns everything that signal can distinguish, progress stalls. The Red Queen Gödel Machine removes that ceiling by evolving the evaluator alongside the agent, so the bar keeps rising as the agent gets better. That is a structural fix to a known bottleneck in recursive self-improvement, not just a tuning trick for one task.

Who it affects

It matters most to researchers and labs building recursive self-improving agents or automated research systems, since the fixed-evaluator ceiling is a shared constraint across that field. The collaboration itself spans Cambridge's Machine Learning Systems Lab, NVIDIA, Flower Labs (a departmental spin-out), MBZUAI and Inria, and the hybrid-model result speaks specifically to anyone weighing open models like NVIDIA's Nemotron against costlier frontier models such as ChatGPT-5.5 for agentic search.

How to use it

There is no product to use yet: the work is a pre-print uploaded to arXiv, and the underlying method has not been released. The team says it will be open sourced to support wider adoption, without giving a release timeline. The closest thing to a usable recipe in the results is architectural: route the bulk of agentic search through a cheaper open model like Nemotron 3 Ultra, and reserve a stronger frontier model like ChatGPT-5.5 for higher-level guidance, which is how the reported 13-times search-token cost cut was achieved.

How solid is it

The claims come from the researchers' own pre-print, not an independent replication or peer-reviewed publication, and this account gives only relative gains (1.78x-1.86x, 9%, about 13 times) without the absolute baseline numbers behind them. Co-author Nic Lane explicitly flagged the cost finding as "a narrow result" that should not be overstated, and the hybrid open-model setup only approached, rather than matched, ChatGPT-5.5's solo performance in the reviewing task tested.

Risks and caveats

The researchers call the work preliminary and say longer search horizons are needed to know how far the approach scales. No benchmark names, dataset sizes or full evaluation protocol are given beyond "scientific paper writing and reviewing" and Math Olympiad-level proof writing and grading, and the method has not been open sourced yet, so outside groups cannot currently verify or extend the results themselves.

“A self-improving agent can only get as good as the test that scores it. The test does not merely measure progress, it defines it, so the efficacy of the test becomes a ceiling the agent cannot climb past.”

— Alex Iacob, PhD student, Cambridge Machine Learning Systems Lab, first author of the paper