Google researchers built a fix for AI agents that cheat their own tests

A team at Google Cloud AI Research has published a method for letting AI agent harnesses rewrite their own tools, prompts, and control logic without quietly memorizing the test they are graded on. The method, called RRSI (Regularized Recursive Self-Improvement), targets a specific failure mode in self-improving systems: a harness that scores well on the benchmark it was evolved against, then loses much of that gain the moment it faces a different task.

That failure mode is the paper’s real subject. According to the researchers, prior self-improvement methods post large score gains on the benchmark split they train on, then see those gains shrink or disappear on tasks outside that split, with two of the compared methods finishing below where they started. RRSI, the authors report, is the only method in their comparison whose score kept climbing once the tasks changed.

The fix is aimed at the evolution loop itself rather than the harness it produces. Each round, a proposer component drafts changes to the harness, working from a shrinking budget that allows several bundled edits early on and forces a single, traceable change later in the run. Every attempt gets logged with what was tried, what it cost, and what it scored, so later rounds build on results instead of retesting ideas that already failed. Two additional checks decide which changes are allowed to stick: a “leakage critic” that throws out any edit that hardcodes task names, answers, or other benchmark-specific shortcuts before it is scored, and a floor that requires a gain to exceed the noise measured on the unmodified harness. A separate rule ties any added inference cost to a matching gain in score, and components that stop paying their way get marked for removal.

Google’s own results, evaluated with Anthropic’s Claude Opus 4.8 as the underlying policy model, show RRSI harnesses improving by an average of 4.0 points on the three benchmarks used during evolution, spanning coding, agentic workplace tasks, and engineering design. On six benchmarks the harness never trained against, average scores rose 3.4 points, and every one of the six improved rather than regressed. The company also reports the resulting harness used 36 percent fewer policy tokens per trial than versions produced by unregularized evolution, which it attributes directly to the cost and pruning rules rather than a side effect.

These are the authors’ own benchmark comparisons, run on evaluation suites the same team assembled, and the paper does not include results from an outside lab replicating the setup. That matters because self-improvement research has a track record of gains that hold up only inside the paper that reports them; RRSI’s own framing, that competing methods overfit their evolve set, is the exact criticism a skeptic would apply to any single team’s numbers.

What RRSI does add to that conversation is a mechanism, not just a claim: it names the specific thing (benchmark-specific edits slipping past the scoring step) that causes the overfitting it says other methods suffer from, and builds a screening step aimed at that mechanism rather than a general regularizer. Teams running their own self-improving coding or agent harnesses now have a documented failure mode to check for: whether their harness’s evolve-set gains survive contact with tasks the loop never saw, and a leakage-style filter is a cheap first thing to add before trusting a benchmark jump.

Based on the RRSI project paper published by Google Cloud AI Research at regularized-rsi.com (arXiv