AI news ·
Google researchers propose method to stop self-improving AI agents from memorizing test answers
RRSI stops self-improving AI agents from memorizing test answers, boosting unseen-task scores by 4.7 points while cutting token use by 30%.
A team from Google Cloud AI Research and several universities has developed a method to stop self-improving AI agents from memorizing their evaluation tests, a problem that can degrade performance when the agent encounters new tasks. The approach, called Regularized Recursive Self-Improvement of Agent Harnesses (RRSI), also cuts token consumption by roughly 30% during runtime.
Self-improving agents iterate on their own code or prompts by running internal tests. Without guardrails, they can overfit to those tests, essentially learning the answers rather than the underlying capability. RRSI introduces three mechanisms to counter this: a budget that caps how many edits a candidate can bundle, a schedule that shrinks that budget over time, and a critic module that rejects changes specific to the benchmark.
Measurable gains on unseen benchmarks
In experiments across eight benchmarks, RRSI raised scores on training tasks by up to 14.1 points. On unseen tasks, the improvement reached 4.7 points. The system achieved these results while using about 30% fewer tokens at runtime than the baseline approach.
The researchers tested the framework with a frozen Claude Opus 4.8 model, and the resulting harness outperformed the baseline on all eight benchmarks. A separate experiment showed that a harness optimized with Gemini 3.5 Flash could transfer gains to a weaker model, lifting Gemini 3.1 Flash Lite accuracy from 11.2 to 14.6. This suggests harness design can benefit model performance even when the underlying model remains unchanged.
Harness design as a safety and efficiency lever
The findings align with parallel work from Nvidia's SoL-Pi and Google's earlier "dream" research, both of which pointed to similar efficiency improvements from better harness construction. The core insight is that how an agent structures its self-improvement loop matters as much as the model inside it.
For teams building agents that write and revise their own prompts, RRSI offers a path to more reliable generalization. Without regularization, an agent that appears to improve during development may simply be memorizing test patterns, creating a false sense of progress. The budget and critic mechanisms provide a straightforward way to reduce that risk while also lowering compute costs.
Why this matters for IT, development, and product teams
For engineers and product developers working with self-improving agent pipelines, the takeaway is practical: overfitting to internal evaluations is a measurable risk, and it can be mitigated with explicit constraints. The 30% token reduction also translates directly to lower API costs. RSI provides a template for building agent scaffolding that generalizes better without requiring changes to the underlying model. Professionals exploring structured approaches to agent design can find relevant coursework in AI Agent Courses and specialized AI R&D Engineering Courses.