微信公众号-PaperWeeklyhttps://www.decemberpei.cyou/rssbox/微信公众号-PaperWeeklyWed, 15 Apr 2026 00:32:46 GMTPyRSS2Gen-1.1.0http://blogs.law.harvard.edu/tech/rssOpenClaw也要装“杀毒软件”?ClawKeeper盯住「龙虾」每一步https://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247719552&idx=1&sn=a88bf66c21ab36eb8ad60266acbfeae9OpenClaw也要装“杀毒软件”?ClawKeeper盯住「龙虾」每一步长推理不一定更强：北航×字节提出SAGE-RL，挖出大模型隐藏天赋https://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247719552&idx=2&sn=e414a3299500b622cc0a8175271890d1长推理不一定更强：北航×字节提出SAGE-RL，挖出大模型隐藏天赋1/3 Token↓，3×性能↑：通义×清华用强化学习重做AgentMemoryhttps://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247719474&idx=1&sn=aa5a0f9eb40a223969f250fbe52e15541/3 Token↓，3×性能↑：通义×清华用强化学习重做AgentMemoryICLR2026|多轮推理越走越偏，T3让Agent找回主动思考https://mp.weixin.qq.com/s?__biz=MzIwMTc4ODE0Mw==&mid=2247719474&idx=2&sn=be225fb932a6dfac03b021ea389476c5ICLR2026|多轮推理越走越偏，T3让Agent找回主动思考