Google Unveils RRSI to Stop Self-Improving AI Agents from Memorizing Benchmarks

In modern artificial intelligence systems, much of the recent performance gains do not stem from updating underlying model weights, but from refining the "harness"—the surrounding scaffolding of prompts, tool integrations, workflows, logic, and memory that dictates an agent's runtime decisions. While automating the rewriting of these harnesses through recursive self-improvement has gained traction, it creates a persistent drawback: agents quickly memorize their evaluation tasks, leading to inflated training scores but poor generalization on unseen problems.

To address this challenge, researchers from Google Cloud AI Research and several collaborating universities have introduced RRSI (Regularized Recursive Self-Improvement of Agent Harnesses). The technique introduces strict guardrails across the self-optimization cycle, ensuring agents generalize effectively to new tasks while lowering compute costs.

The Overfitting Trap in Self-Refining Harnesses

A harness acts as the operational brain around a frozen large language model, guiding whether the agent examines the correct files before modifying them, recovers gracefully from errors, and delivers structured outputs. Traditionally, engineers spent hours manually diagnosing failed trajectories and patching harness instructions.

Recent recursive self-improvement pipelines automate this loop by instructing an LLM to rewrite its own harness based on performance feedback. However, repeatedly tuning against a fixed suite of test benchmarks causes the system to overfit. Agents develop search patterns specific to a single benchmark, reward candidates that succeed purely by luck, and accumulate unnecessary complexity that boosts benchmark numbers without imparting real-world capability. On novel tasks, performance often stagnates or drops below the original unoptimized baseline.

Regularization via Edit Budgets and a Strict Critic

RRSI intervenes at two critical junctions in the optimization loop while keeping the harness fully editable: when changes are proposed and when they are accepted.

To control the proposal phase, the system enforces a shrinking "edit budget." Early optimization rounds permit broader, systemic rewrites of the harness. As rounds progress, the budget contracts, restricting the system to smaller, modular edits whose impacts can be cleanly traced. RRSI also maintains a history of past iterations to prevent retrying failed strategies, and if improvement plateaus, it deliberately explores unedited segments of the harness.

Before any proposed modification is merged, a dedicated critic model evaluates the code. The critic filters out proposals that hardcode task names, benchmark-specific workarounds, or leaked solutions. Furthermore, RRSI enforces an efficiency rule: changes that introduce higher compute overhead are rejected unless accompanied by measurable accuracy gains, and obsolete, unhelpful components are pruned.

Generalization Across Unseen Benchmarks

The research team evaluated RRSI against an unmodified baseline and four existing automated optimization methods across eight benchmarks covering engineering design, agentic office environments, and software coding. The underlying model, Claude Opus 4.8, remained frozen throughout testing.

The evaluation revealed distinct advantages for the regularized approach:

• Sustained Generalization: RRSI achieved score increases of up to 14.1 points on training tasks and gains of up to 4.7 points across five unseen benchmarks, led by a 4.7-point boost on JobBench.
• No Regression Below Baseline: Unlike competing methods—two of which degraded below the initial baseline harness on unseen benchmarks—RRSI maintained positive gains across all unencountered tasks.
• Compute Efficiency: The resulting harnesses used approximately 30 percent fewer runtime tokens than unregularized alternatives.

Although RRSI yielded smaller training benchmark gains than unconstrained methods, this trade-off directly prevented the memorization that crippled other approaches.

Cross-Model Portability

The researchers discovered that harnesses refined through this process also benefit less capable models. A software engineering harness initially optimized using Gemini 3.5 Flash improved the accuracy of Gemini 3.1 Flash Lite from 11.2 to 14.6 points without requiring any task-specific adaptations. This suggests that the structural workflows discovered by the system translate into generalized operational improvements rather than model-exclusive quirks.

The authors noted that the current study focuses solely on harnesses wrapping frozen foundational models, rather than scenarios involving direct weight updates. The code for the RRSI framework has been made publicly available on GitHub.