#RRSI
Google 搞了个叫 RRSI 的正则化方法，专治自改进 Agent 背题，新基准上多拿 4.7 分还省三成 token。
This framework represents a significant shift in how AI agents can autonomously optimize their components while maintaining model stability. Developers working on complex AI agent systems can leverage RRSI to create more adaptable and efficient agents without constant manual retraining, potentially reducing development time and improving agent performance in real-world applications.
Optimizing an agent's harness against a benchmark can inflate eval scores while degrading real-world performance, and this paper quantifies that gap with held-out numbers. Builders running automated prompt or harness optimization should add critic and pruning loops like RRSI's to protect out-of-distribution performance. The paper and a chat-with-paper interface are available via arXiv and dair.ai.
Regularization as the safety valve for recursive self-improvement suggests the main failure mode of self-modifying agents is instability rather than capability. Researchers building agent frameworks and self-improving loops can study the constraints as a template for controlled iteration. The work appears on Hugging Face Papers, making it directly comparable with other open agent-harness research.
Google开源了RRSI：给自动进化智能体harness的方法加正则化，分布外基准还能涨4.7分，还省30%的token。