AIThe Decoder2h ago
Google researchers find a way to keep self-improving AI agents
Google researchers find a way to keep self-improving AI agents from memorizing their tests
TL;DRGoogle's new method stops AI from cheating on tests by memorizing them instead of learning.
Why it matters: Better generalization means AI systems that actually improve rather than just gaming benchmarks.
Self-improving AI agents tend to memorize their test tasks, so their gains shrink or disappear on new ones. RRSI, a new method from Google researchers, reins in this effect and lifts scores on unseen benchmarks by up to 4.7 points while using about 30 percent fewer tokens than…
Read full articleSource: The Decoder · Opens in new tab