The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators
arXiv paper · source date 2026-06-24 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog
1
Problems / challenges / motivations
- Self-improving agents are bounded by the evaluator that scores their modifications, so a fixed evaluator caps improvement.
- In hard-to-verify domains — paper writing, proof grading — there is no benchmark to serve as the evaluator at all.
- Letting the evaluator evolve freely invites the agent and evaluator to drift together into a shared delusion.
2
Key ideas
- A recursive self-improvement framework where learned evaluators co-evolve with the agents they score.
- "Controlled utility evolution" is the safeguard: the evaluator is frozen within an epoch, and a challenger evaluator replaces the incumbent only if it beats it on held-out human ground truth.
- Extends the Darwin/Huxley Gödel lineage into hard-to-verify domains while retaining per-epoch improvement guarantees.
- Cambridge, NVIDIA, MBZUAI, Inria, and Flower Labs.
3
Why it matters for AI engineering
- The first RSI system whose evaluator evolves alongside the agent, addressing the evaluator-reliability ceiling directly rather than working around it.
- The freeze-within-epoch plus human-ground-truth-gated replacement is a transferable pattern for anyone letting a judge change over time.
- Caveat: a fresh late-June preprint, largely unreplicated. The per-epoch guarantees depend entirely on the quality of the held-out human ground truth, and co-evolving evaluators risk the recursive-drift failure modes SkillLearnBench and MemoryArena document.
Comments
No comments yet.