Google Introduces RRSI Method to Prevent Agents from Memorizing Test Tasks

Summary

Google researchers propose RRSI, an approach that regulates the self-reflective self-improvement process of agent scaffolds. By employing shrinking regulation budgets and a strict critic, the system prevents agents from memorizing benchmark tasks. This method sacrifices some gains on the training set to achieve meaningful generalization to unseen tasks.

Why it matters

The industry is shifting from manually tuned agent scaffolds to large language model (LLM)-driven self-optimization loops. The memorization issue identified in this study is a real and underappreciated failure mode that can silently inflate benchmark scores without improving actual competence. RRSI offers a concrete, open-source protection mechanism that other teams can adopt. Furthermore, the ability to transfer scaffold improvements across model tiers (from Gemini Flash to Flash Lite) points toward a path for building cheaper and more robust agent stacks.

Highlights

• Shrinking regulation budgets combined with a critic that rejects hardcoded task names/solutions form the core anti-memorization mechanism.
• RRSI achieved the lowest gain on the training set across all variants. However, it was the only method to significantly outperform the baseline on unseen benchmarks (up to +4.7 points).
• A scaffold optimized with Gemini 3.5 Flash improved a weaker Gemini 3.1 Flash Lite model from 11.2 to 14.6 points with zero modifications.

Critical take

The 'self-reflective self-improvement' framework overstates a system that is actually a rule-based regulation loop. The critic's rejection criteria (no hardcoded task names, no benchmark-specific cheating) rely on human-written heuristics. Consequently, the system is far from autonomous self-improvement. Moreover, the absolute gains on unseen tasks (≤4.7 points) are limited. The study is limited to frozen models, and four comparison baselines are labeled as unnamed 'recent methods.' This makes it difficult to assess whether RRSI's superiority is genuinely significant or merely a result of weak competitors.