Google research proposes RRSI, preventing self-improving agents from memorizing test tasks
Google researchers propose RRSI (Regularized Recursive Self-Improvement of Agent Harnesses), using a gradually shrinking edit budget and strict critic review to suppress language models' memorization of test tasks when repeatedly rewriting agent harnesses.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original