Google Stops Agents From Cheating Tests

🔥 What happened Google Cloud AI Research and several universities built RRSI, a method that stops self-improving AI agents from memorizing their test tasks. It caps how many edits a candidate harness can bundle and uses a critic to reject benchmark-specific tricks. 💡 Why it matters Prior optimization methods lose up to 4.7 points on unseen benchmarks or even drop below baseline. RRSI gains 4.7 points there and cuts runtime tokens by 30 percent. A coding harness tuned with Gemini 3.5 Flash lifted a weaker model from 11.2 to 14.6 points. ⚡ Our take Finally someone treats agent overfitting as the real problem it is. Tuning your agent on the test set doesn't build a better agent, it builds a better cheater.