Google Researchers Discover Method to Stop Self-Improving AI Agents from Memorizing Test Questions

Meta description: Google RRSI gains up to 14.1 points on training tasks and 4.7 points on unseen benchmarks, while agent runtime token use falls about 30 percent.

Tags: RRSI, agent harnesses, Claude Opus 4.8, Gemini 3.5 Flash, JobBench

Featured image: An abstract technology visual of feedback loops, branching edit paths, and shrinking budgets, rendered with no people, company logos, or text.

A new research paper on RRSI says much recent progress in AI agents comes from agent harnesses, not new models. A harness is the control layer around an agent: it determines whether the agent opens the correct file before an edit, recovers after an error, and presents the outcome cleanly. This matters because transferable capability can be hidden by benchmark memory.

From Manual Patching to Automated Rewrites

Before the newest methods, engineers inspected failed runs and repaired the harness directly. Newer systems close the loop automatically: a language model edits the harness over repeated rounds, guided by signals from test tasks. The paper frames this as practical recursive self-improvement. Feedback generated by the system updates the harness, and the updated harness governs later system behavior.

The Memorization Problem

When the same small pool of test tasks remains in the loop, the agent can learn the tasks instead of the skills. Training scores climb, but unseen-task gains may shrink or vanish. The paper describes several failure modes: the optimizer can lock onto benchmark-specific quirks, choose edits that happen to score well, or add complexity that improves a metric without improving agent behavior. Other methods mostly improve training tasks, and little carries over to unseen benchmarks. Google-credited figures in the paper show RRSI raising scores in all three domains: coding, agentic office work, and engineering design.

Edit Budgets, the Critic, and the 14.1 Tradeoff

RRSI, or Regularized Recursive Self-Improvement of Agent Harnesses, keeps the harness fully editable but tightens both the proposal side and the acceptance side of the loop. A candidate can bundle only a limited number of independent edits, controlled by a shrinking budget. Large rewrites are allowed early; later rounds require small changes that can be clearly traced to a result. The system stores earlier attempts to avoid repeating failed ideas and, if gains plateau, probes untouched harness regions. Google-credited figures label this two-point control, since RRSI reins in self-optimization when changes are proposed and when they become permanent.

During selection, a critic rejects proposals that hardcode task names, solutions, or other benchmark-specific tricks. Extra compute is accepted only when it yields a measurable gain, and stale components are pruned.

The evaluation used eight benchmarks across coding, agentic office work, and engineering design, with Claude Opus 4.8 frozen throughout. Compared with the unmodified baseline harness and four recent optimization methods, the paper reports up to 14.1 points on trained tasks and up to 4.7 points on five unseen benchmarks. Runtime token use is about 30 percent lower than the unregularized version. On unseen benchmarks, performance never dropped below the baseline, a common failure for memorized harnesses. It improved on every outside-training task, with the largest gain, 4.7 points, on JobBench.

All methods scored well on training tasks, but on new tasks two fell below the baseline harness. RRSI had the smallest training gain among variants, yet was the only method clearly above baseline on unseen tasks; the guardrails aim to produce exactly this tradeoff. Among optimized harnesses, RRSI used the fewest tokens and steps and was strongest on new tasks, though the unmodified baseline was leaner.

Cross-Model Transfer and the Frozen-Model Boundary

Transfer was not limited to the model used for optimization. A coding harness optimized with Gemini 3.5 Flash lifted the weaker Gemini 3.1 Flash Lite from 11.2 to 14.6 points with no model modification. The discovered mechanisms did not depend on the discovering model's capability. The study covers only harnesses around frozen models, not changing model weights. The authors conclude self-improvement is reliable only when repeated feedback becomes lasting changes. The code is available on GitHub.

The current stage sits between manual tuning and model training. Manual harnesses often fail to generalize: ARC-AGI-3 tests showed Opus 4.6 with a purpose-built harness scored 97.1 percent in a familiar environment and 0 percent in an unfamiliar one. Nvidia recently introduced SoL-Pi, a related method where a research agent automatically rebuilds coding-agent harnesses and cuts token use by up to 49 percent without noticeable performance loss. Shortly before that, Google made agents "dream" about past search runs to improve search strategy, again leaving the model unchanged. The question is how far harness-level self-improvement can go before it becomes benchmark memory rather than transferable engineering.