The Red Queen Has a Patent: Why a Cambridge–NVIDIA Preprint Just Reset – Good Machine

TL;DR

A 13-author preprint from the University of Cambridge, NVIDIA, MBZUAI (Abu Dhabi), Inria (France), and Flower Labs — published to arXiv on 24 June 2026, viral on 28 June — describes the first recursive self-improvement system whose evaluator evolves alongside the agent it grades.
The mechanism, called controlled utility evolution, divides the search into epochs with frozen evaluators and lets the judge be replaced only when a challenger beats the incumbent on a held-out human ground-truth set.
Concrete numbers, not vibes: 1.35×–1.72× fewer search tokens on Polyglot coding; co-evolved paper-writers reach 1.78×–1.86× higher acceptance rates; co-evolved proof-graders are 9% more accurate on Olympiad-level proofs. The system also corrects a known AI-judge bias — baseline reviewers over-accepting AI-written papers at 1.91× the human rate — by breeding adversarial evaluators.
The paper lands three weeks after Jack Clark (Anthropic) put a 60% probability on recursive self-improvement arriving by end-of-2028. The RQGM is not that system. It is the first credible mechanism that points toward it.
What this isn't: a deployed product, a peer-reviewed result, or proof of intelligence explosion. It is a research preprint with sharp numbers and a publicly stated limitation — the authors say so themselves.

What happened

On 24 June, Alex Iacob, Andrej Jovanović, and William F. Shen — joint first authors at Cambridge — with eleven co-authors spanning Cambridge's Machine Learning Systems Lab, NVIDIA, MBZUAI in Abu Dhabi, Inria in France, and Flower Labs. Title: The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators. Thirty-seven pages. One central idea.

For four days, the paper sat where most arXiv preprints sit — visible to specialists, invisible to anyone else. Then on 28 June it broke containment. Tech Times ran a long technical write-up framing it as "the second-order alignment question." Chinese-language AI media — Xinzhiyuan, picked up by KuCoin — called it "the most dangerous AI paper of the year." Inside AI X/Twitter, the citation chain rapidly fused with Jack Clark's May post putting a 60% probability on recursive self-improvement by end-of-2028.

That fusion is the story. The Clark prediction has been sitting in the discourse for seven weeks looking for a mechanism that would either validate or deflate it. The Red Queen Gödel Machine is the first paper anyone has been able to point to and say: this is what part of that mechanism might look like.

The framework first: what self-improving AI systems were missing

To see what the RQGM changes, you have to see what every prior self-improving system shared.

The theoretical ancestor is Jürgen Schmidhuber's 2003 Gödel machine: a hypothetical program that would rewrite its own code only after producing a formal mathematical proof the rewrite was an improvement. The proof requirement turned out to be computationally intractable. The idea sat in storage for two decades.

Practical descendants — the Darwin Gödel Machine (2024) and the Huxley Gödel Machine — replaced the proof requirement with evolutionary search: spawn mutant code variants, sandbox them, discard failures, keep survivors. These systems posted real gains on agentic coding benchmarks. But they shared a single architectural constraint: the evaluator was fixed. Every variant was graded against a benchmark or judge that did not change.

This is a structural version of Goodhart's Law — any measure becomes corrupted when treated as an optimization target. A self-improving system pressing against a frozen evaluator is, by definition, an optimizer whose incentive is to game the evaluator. The longer the system runs, the less the score means.

The biologists' Red Queen Hypothesis — proposed by Leigh Van Valen in 1973 to explain why species must constantly evolve simply to maintain fitness relative to other evolving species — gives the paper its name and its claim: a static evaluator inside a co-evolutionary system is, functionally, extinct.

The mechanism: controlled utility evolution

The RQGM's contribution is a specific algorithm for letting the evaluator evolve without making the search unstable. The authors call it controlled utility evolution, and the design is more conservative than the headlines suggest.

The search is divided into discrete epochs. Within an epoch:

The evaluator is frozen. It scores every agent variant. Prior self-improvement guarantees from the Darwin/Huxley Gödel lineage apply directly, because the optimization target is stable.
In parallel, the system co-evolves challenger evaluators in a shared codebase.

At the epoch boundary — and only at the boundary — a challenger evaluator can replace the incumbent. The replacement is gated by a strict rule: the challenger must demonstrate, on a held-out human-curated ground-truth dataset, that it statistically outperforms the current evaluator. If it doesn't, the incumbent stays.

When replacement does happen, the system performs selective erasure: it discards only the historical scores that came from the deposed evaluator. Everything else — the agent archive, the search policy, the substrate — survives. A new epoch begins.

If that sounds like a workable answer to the Goodhart problem, it is — at one level of abstraction. The catch is what the authors say themselves, in the paper's own words: the system "loosens convergence guarantees compared to static evaluation criteria." Translated: the ground-truth dataset itself is now the static benchmark. The same Goodhart pressure that the RQGM defuses at the agent–evaluator layer reappears, one level up, between the evaluator and the ground-truth anchors. It just takes longer to bite.

What the numbers actually say

The empirical section is where the paper earns or loses its citations. Read carefully, it earns them — with caveats.

Verifiable coding (Polyglot benchmark):

RQGM reaches a held-out pass rate of 71.7% versus the prior SOTA's 69.9%.
It does so using 1.35×–1.72× fewer search tokens than HGM-H, because the co-evolved code reviewer is queried once rather than running multi-turn test execution.

Scientific paper writing (agent-as-judge panel):

Co-evolved writers raise acceptance rate from a SOTA baseline of 21.8% to 40.5% — a 1.78×–1.86× lift.

Olympiad-level proof grading:

Co-evolved graders reach 9% higher ground-truth accuracy than static baselines, at roughly 3× lower search cost than the prior SOTA.

Bias correction (the one most people will quote):

The strongest baseline reviewer over-accepts AI-generated papers at 1.91× the human rate — a known failure mode in agent-as-judge evaluation.
RQGM corrects this by harvesting adversarial AI submissions at each epoch boundary and rewarding new evaluators that correctly reject them. After several rounds, the resulting judge treats AI and human papers with equivalent stringency while keeping 80% accuracy on ground-truth validation.

The bias-correction result is the most genuinely surprising part of the paper. It says the same machinery that makes self-improvement faster can also make a class of well-documented AI evaluation failures go away. That is not a trivial finding.

Where this sits in the recursive self-improvement debate

Here is where I want to be specific about what this paper is and isn't, because the discourse is already collapsing the distinction.

What the RQGM is: a small-scale empirical demonstration of a previously theoretical principle — that an agent and its evaluator can be co-evolved under formal stability constraints, and that doing so produces measurable gains on three distinct benchmarks. The system runs as a research artefact. The authors call it a "preliminary empirical investigation." Peer review has not happened. The longest experimental run is bounded; the paper notes that longer search horizons are future work.

What the RQGM is not: an autonomous AI improving itself in the wild, an intelligence explosion proof of concept, or evidence that Jack Clark's 60%-by-2028 number is correct. It is — and this is the honest framing — the first piece of public engineering evidence that gives the Clark prediction a real mechanism to argue about.

Until 24 June, the recursive self-improvement debate was conducted in two registers: confident probabilistic forecasts from people inside frontier labs (Clark at 60%, Gary Marcus at under 10%) and toy-scale demonstrations on GPT-2-class models. The RQGM, together with Amazon's A-Evolve paper from 9 June (the autonomous post-training run on a 30B-parameter NVIDIA Nemotron model that placed 8th out of 4,000 leaderboard entries), are the first two papers that bring the debate inside the empirical regime where it can actually be tested.

The two papers point in slightly different directions and that matters. A-Evolve showed an autonomous loop can survive contact with frontier-scale execution and self-correct a broken proxy metric mid-run. The RQGM shows the loop can co-evolve its own grading criteria with formal stability guarantees per epoch. Combine them and you have something close to a blueprint — one that nobody, yet, has run end-to-end.

Stakeholder landscape: who benefits from this being everywhere

Real beneficiaries:

Cambridge ML Systems Lab and Flower Labs — concrete reputational and recruiting gains. Nicholas Lane's group just shipped one of the year's most-cited preprints, with NVIDIA and MBZUAI co-authorship as institutional cover.
NVIDIA Research — second high-visibility self-improvement paper in a fortnight (A-Evolve on 9 June, RQGM on 24 June). The narrative is no longer "NVIDIA sells the picks and shovels"; it is "NVIDIA's research arm is publishing the recursive self-improvement engineering."
MBZUAI — Abu Dhabi's research-credibility play continues. UAE on an RSI paper alongside Cambridge and NVIDIA is exactly the institutional positioning the country has been buying for three years.
The AI safety policy community — gets a citable, concrete artefact to point to in regulatory hearings. The February 2026 International AI Safety Report flagged RSI infrastructure as a cross-cutting national-security risk. The RQGM gives that flag teeth.

Real losers:

Anyone whose business model assumed static AI evaluation was a defensible moat. Benchmark vendors, eval-as-a-service tooling, agent evaluation startups — the RQGM does not kill these markets but it does shift the technical baseline they have to argue against.
Skeptics of near-term RSI. Marcus's sub-10% number now has to absorb a concrete result it didn't have to absorb a fortnight ago. That doesn't make Clark right. It does make the conversation harder.

Who's loud but actually peripheral:

The accounts framing this as "ASI by 2028 confirmed." It isn't. The paper does not claim that, and reading the numbers carefully makes the leap untenable.
The accounts framing it as "another arXiv paper, nothing happened." That isn't right either. The mechanism is genuinely new and the benchmark deltas are not noise.

Cross-layer implications

For AI infrastructure: the RQGM uses 1.35×–1.72× fewer tokens than the prior SOTA on coding tasks. If that ratio holds as systems scale, it changes the unit economics of agentic coding evaluation. Eval cost has been the silent line item in frontier-lab budgets for two years; mechanisms like this are how it stops being silent.

For AI safety policy: the held-out ground-truth dataset is the safeguard. It is also, structurally, a static benchmark at one level higher than the one the RQGM was designed to escape. The next paper — and there will be a next paper, within months — will likely propose evolving the ground-truth set itself. When that paper lands, the policy question stops being theoretical.

For Asia-Pacific research positioning: the author list — Cambridge (UK) lead, NVIDIA (US), MBZUAI (UAE), Inria (France), Flower Labs (UK-incorporated, distributed) — is the most geographically distributed frontier RSI paper to date. The US-China binary that dominates the regulatory discourse does not capture where the frontier work is actually being done.

For the OpenAI/Anthropic release-gate story breaking in parallel: GPT-5.6 Sol entered limited-preview release on 26 June at the request of the Trump administration, citing cyber-capability concerns. Anthropic's Mythos 5 was cleared by the US government on 27 June for vetted US organizations only; Fable 5 remains restricted. Read alongside the RQGM, the regulatory moves look less like ad-hoc political theatre and more like Washington reading the same papers AI Twitter is reading and trying to install brakes before the next set of mechanisms compound.

Re-aim: what this means for different readers

For the general public: nothing in this paper changes your week. It is research, not deployment. The headlines calling it "the most dangerous AI paper of the year" are doing what headlines do — competing for attention. The honest framing is that a credible group of researchers showed a previously theoretical idea works on small-scale problems, and the result will be cited by both safety advocates and capability boosters for the rest of the year. Pay attention to what your government's AI safety institute says about it over the next quarter, not to what gets posted on X today.

For practitioners building on agentic coding stacks (Codex, Claude Code, Kimi Code, Cursor Agent, Devin, and similar): two specific things to track. First, the agent-as-a-judge code-review signal the RQGM uses is the kind of architectural pattern that will start appearing in production agent frameworks within months — it materially reduces eval cost and improves pass rates. Second, if you have benchmark-validation steps in your CI/CD that assume a fixed eval, the RQGM is the signal that this assumption has a shelf life. Start thinking about how your eval infrastructure handles non-stationary judges before someone ships one to you.

For AI safety researchers and policy professionals: the paper is doing your work for you. It explicitly names the second-order Goodhart problem and explicitly states that loosening convergence guarantees is the cost. If you are preparing submissions to the EU AI Office, the UK AI Safety Institute, or any equivalent body, the RQGM is now the citation that gives "recursive self-improvement" a defined mechanism rather than a vibe. The mechanism is the part that lets regulation grip.

For investors and analysts covering AI infrastructure: the eval-cost line in frontier-lab cost structures is about to come under sustained pressure. Companies whose AI evaluation businesses assume static benchmarks remain the dominant evaluation paradigm should be re-stressed against a scenario where co-evolutionary evaluation becomes the engineering default in 18 months.

For everyone: there is no urgent personal action. If your AI coverage diet is making you anxious about ASI in 2028, this paper is not the reason to escalate that anxiety. It is, however, a good reason to keep paying attention.

Uncertainty ledger

Peer review: not complete. The paper is a preprint. Replication by independent groups has not been published.
Scale: the experiments run on bounded search horizons. The authors explicitly flag longer-horizon behaviour as future work.
Generalization: results are reported on three task families (coding, paper writing/reviewing, proof writing/grading). Whether the mechanism transfers to open-ended scientific research — the obvious target — is not yet established.
The second-order Goodhart problem: the paper names it; it does not solve it. The held-out ground-truth dataset is a stationary anchor and a sufficiently capable system pressing against the epoch-boundary replacement criterion could eventually exploit it. The authors do not claim otherwise.
What would change the analysis: a follow-up paper extending the safeguard to a non-stationary ground-truth set, a peer-reviewed replication at larger scale, or an independent benchmark of the bias-correction result on a broader range of evaluator types.

Bottom Line

The Red Queen Gödel Machine is not an autonomous AI building itself, and the people calling it that are not reading the paper. It is the first credible piece of engineering evidence that the recursive self-improvement debate now has a working mechanism to argue over — and the mechanism's central safeguard is exactly the thing the next paper will be designed to dissolve. Treat this as the start of a research direction, not the end of one. The interval between credible RSI papers just collapsed from years to weeks; that is the actual news.

Sources

Tier 1
- Iacob, A., Jovanović, A., Shen, W. F., et al. The Red Queen Gödel Machine: Co-Evolving Agents and Their Evaluators. arXiv:2606.26294v1, 24 June 2026. (Primary document.)
- Reuters, Austria lobbies EU to host Anthropic AI after US curbs, 28 June 2026 (cross-context).
- TechCrunch, OpenAI limits GPT-5.6 rollout after government request, 26 June 2026 (cross-context).
- The Guardian, OpenAI staggers AI model release after Trump administration request, 26 June 2026.
- The Verge, OpenAI unveils GPT-5.6 amid US AI regulatory drama, 26 June 2026.
- Cambridge Machine Learning Systems Lab (mlsys.cst.cam.ac.uk) — institutional confirmation of authorship.
Tier 2
- Tech Times (Jerry Owens), Recursive Self-Improvement Now Has a Co-Evolving Evaluator: Cambridge–NVIDIA Paper Raises the Stakes, 28 June 2026.
- Tech Times (Richard L. Wells), NVIDIA AI Trained Itself on a 30B Model (A-Evolve coverage), 26 June 2026.
- Forbes, OpenAI Limits New AI Models To 'Trusted Partners' Vetted By U. S. Government, 26 June 2026.
- Bloomberg via Reuters, Austria Lobbies EU to Host Anthropic After US Access Curbs, 28 June 2026.
Tier 3
- KuCoin / Xinzhiyuan (Chinese-language coverage), NVIDIA paper on AI self-evolution sparks debate over 2028 ASI timeline, 28 June 2026 (virality signal in Asia-Pacific discourse).
- The Neuron newsletter, daily AI digest, 28 June 2026 (Western practitioner virality signal).
- Jack Clark (Anthropic), public X post on 60% RSI-by-2028 probability, 4 May 2026 (referenced context, not load-bearing).