Apple 发布 SCLATE：持续学习智能体训练与评估基座
原标题：SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation
AI 摘要
Apple 研究团队发布 SCLATE，一个用于持续学习智能体训练与评估的执行基座。它通过适配器让基准测试和未修改的智能体将事件加入统一调度器，并用混合模拟时钟把长达一个月的场景压缩到数小时。团队移植了七个基准，在十个模型上对比十种未修改的 harness 与记忆配置，发现额外记忆系统未必优于 harness 原生记忆。随后用未修改的 harness 和记忆系统后训练 Qwen3.5-4B，读取文件行数减少 6.8 倍，SWE-bench Verified 通过率提升 16.7 个百分点，MetaClaw 准确率最高提升 11.8 个百分点。
正文节选
SCLATE: A Substrate for Continual-Learning Agent Training and Evaluation AuthorsYoungmok Jung, Sirajul Salekin, Henry Tran, Javier Movellan, Zhao Huang, Manjot Bilkhu Continual-learning agents are systems of models, harnesses, and memory operating over long multi-session horizons. Evaluating and training them requires interleaving tasks with agent-side events such as session stop and start, crons, and memory consolidation. Yet existing benchmarks and training frameworks schedule only the benchma