DefensesarXiv 2609.35561·Sep 29RSI-Master:用结构化实验引导模型自主改进提出 RSI-Master,约束自主后训练实验动作并结构化探索以防奖励作弊arXiv2609.35561PublishedSep 29LayerApplication layerSettingscience agent, Multi-agent systemsDefensesPermissions and approvalRiskReward hackingComponentPermissions and sandboxingPaper analysisRead analysis