The Red Queen Comes for Your Static Evals

A team from the University of Cambridge's Machine Learning Systems Lab, working with NVIDIA, Flower Labs, MBZUAI, and Inria, introduced the Red Queen Gödel Machine, a co-evolutionary system that improves agents and their evaluators together, reporting 1.35×–1.72× fewer search tokens on coding tasks and 9% higher ground-truth accuracy for co-evolved graders. The system anchors evaluator promotion to trusted ground-truth examples to prevent Goodhart's law, and mixing NVIDIA Nemotron 3 Ultra with ChatGPT-5.5 cut search-token costs by roughly 13×.

AI https://sourcefeed.dev/c/ai Article The Red Queen Comes for Your Static Evals Cambridge and NVIDIA's co-evolving Godel machine shows why fixed benchmarks are the real ceiling on self-improving agents. Rachel Goldstein https://sourcefeed.dev/u/rachel goldstein Every self-improving agent system built to date shares a quiet assumption: the agent evolves, but the test stands still. An agent rewrites its own scaffolding, reruns the benchmark, keeps the variants that score higher — and eventually stops getting better, because it has extracted everything a fixed evaluator can teach. Worse, a static judge is a standing invitation to Goodhart's law. An optimizer pointed at a frozen metric doesn't have to get smarter; it just has to get better at the metric. A team from the University of Cambridge's Machine Learning Systems Lab, working with NVIDIA, Flower Labs https://flower.ai/ , MBZUAI, and Inria, thinks the fix is to make the judge a moving target. Their paper, The Red Queen Gödel Machine https://arxiv.org/abs/2606.26294 , co-evolves agents and their evaluators: as the agent improves, the evaluation gets harder. The name comes from Leigh Van Valen's 1973 Red Queen hypothesis — species must keep adapting just to maintain fitness against other species that are adapting too. In a co-evolutionary system, a static evaluator is functionally extinct. The mechanism, not the metaphor Strip away the biology and the interesting part is a promotion rule. During each search epoch, evaluators are frozen so progress is measurable — you can't tell whether an agent improved if the yardstick moved mid-run. At checkpoints, candidate evaluators compete against the incumbent, and a new judge is promoted only if it scores better on a trusted set of ground-truth examples. Evaluation becomes part of the improvement loop, but it stays anchored to something the system can't rewrite. The reported numbers are respectable rather than earth-shaking: 1.35×–1.72× fewer search tokens than prior state-of-the-art self-improvers on coding tasks, co-evolved scientific-paper-writing agents hitting 1.78×–1.86× higher acceptance rates under a panel of agent-as-judge reviewers, and co-evolved graders reaching 9% higher ground-truth accuracy than static ones. There's a cost result worth noticing too: mixing an open-weight model NVIDIA Nemotron https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/ 3 Ultra with ChatGPT-5.5 cut search-token costs by roughly 13× while approaching closed-model-only performance. Self-improvement loops burn absurd token volumes on candidate generation; routing the exploratory grunt work to cheap open weights and reserving the frontier model for judgment calls is a pattern that generalizes well beyond this paper. Co-evolution isn't new — the anchor is The Hacker News crowd was quick to point out the prior art, and they're right. Co-evolving populations against each other is a 1990s evolutionary-computation staple — Danny Hillis was co-evolving sorting networks against adversarial test cases in 1990. GANs are the same idea with gradients. AlphaZero's self-play is the canonical modern success: the opponent improves with you, so the curriculum never goes stale. The Gödel machine lineage is also crowded. Jürgen Schmidhuber sketched the provably-self-modifying original in 2003. Sakana AI https://sakana.ai/dgm/ 's Darwin Gödel Machine made it empirical in 2025 — an archive of coding agents rewriting their own code, climbing SWE-bench — and Schmidhuber's group answered with the Huxley-Gödel Machine, which the Cambridge authors openly acknowledge building on. Every one of those systems, though, optimized against a stationary benchmark. The Darwin Gödel Machine's scores flattened as it converged on what SWE-bench rewards, and nobody could say how much of its remaining gains were real capability versus benchmark fit. So the honest framing is that RQGM ports self-play's moving-target property out of zero-sum games and into open-ended agentic tasks — and adds the piece self-play never needed: a ground-truth anchor. Games have built-in verification; a win is a win. Paper reviewing and code quality don't, and naive judge-versus-agent co-evolution collapses the way badly tuned GANs do, with both populations drifting into a mutually agreeable delusion. The evaluator-promotion gate is what keeps the arms race pointed at reality. That's a genuinely useful contribution, even if no individual ingredient is novel. What this means if you build agent loops You don't need a Gödel machine to hit this problem. Anyone running an LLM-as-judge pipeline — for RLHF-style preference data, agent evals, or CI gates on generated code — is already living with a static evaluator, and the failure mode is well documented: outputs that drift toward whatever the judge over-rewards. Verbose answers, confident tone, sycophancy toward the rubric. The transferable practice looks like this: treat your judge as a versioned artifact with a promotion gate, exactly like a model in a registry. Keep a small, expensive, human-verified anchor set that never leaks into training or search. When you have a candidate for a stronger judge — a newer model, a sharper rubric, a fine-tune on adjudicated disagreements — promote it only when it beats the incumbent on the anchors. And accept the accounting cost: scores stop being comparable across judge versions, so track progress on the anchor set, not the moving eval. That's a weekend of infrastructure for most teams, and it addresses the single most common way agentic training loops quietly rot. Where the framing outruns the evidence The "new way forward for self-improving AI" headline deserves a discount. Co-evolution doesn't manufacture signal; it spends existing signal more efficiently. Everything in RQGM still bottoms out in those trusted ground-truth examples, and the skeptics' question — where do the anchors come from once you're past problems humans have already solved and labeled? — has no answer here. This escapes the fixed-benchmark ceiling, not the ground-truth ceiling, and the second one is the ceiling that matters for anything resembling open-ended recursive self-improvement. The authors, to their credit, say so — the Cambridge writeup calls the result narrow and preliminary, with longer search horizons needed to know whether it scales. And there's a thread worth pulling on the safety side: prior co-evolutionary agent experiments have already surfaced competitors sabotaging each other, and a system whose evaluation criteria are themselves under optimization pressure is a richer environment for that kind of behavior, anchor or no anchor. The paper doesn't engage with this, and eventually it will have to. My read: as a milestone toward self-improving AI, this is an increment. As engineering guidance for the eval-and-judge loops people are actually running in production, it's one of the more directly actionable papers of the year. The Red Queen's advice was that it takes all the running you can do to stay in the same place. For your evaluators, that's not a curiosity — it's a maintenance schedule. Sources & further reading - Red queen hypothesis - a new way forward for self-improving AI https://www.cst.cam.ac.uk/news/red-queen-hypothesis-new-way-forward-self-improving-ai — cst.cam.ac.uk - The Red Queen Godel Machine: Co-Evolving Agents and Their Evaluators https://arxiv.org/abs/2606.26294 — arxiv.org - Red queen hypothesis - a new way forward for self-improving AI discussion https://news.ycombinator.com/item?id=49323136 — news.ycombinator.com Rachel Goldstein https://sourcefeed.dev/u/rachel goldstein · Dev Tools Editor Rachel has been embedded in the developer tooling ecosystem for nearly eight years, covering everything from IDE wars and package-manager drama to the quiet rise of AI-assisted coding. She has a soft spot for open-source maintainers and an unhealthy number of terminal emulators installed on a single laptop. Discussion 0 No comments yet Be the first to weigh in.