文章预览
LG - 机器学习 CV - 计算机视觉 CL - 计算与语言 AS - 音频与语音 RO - 机器人 1、[LG] Can Looped Transformers Learn to Implement Multi-step Gradient Descent for In-context Learning? 2、[LG] Don't Transform the Code， Code the Transforms：Towards Precise Code Rewriting using LLMs 3、[CL] Agents Thinking Fast and Slow：A Talker-Reasoner Architecture 4、[LG] IGNN-Solver：A Graph Neural Solver for Implicit Graph Neural Networks 5、[LG] Language model developers should report train-test overlap 摘要：环形Transformer能否学会在上下文学习中实施多步梯度下降、用LLM实现精确代码重写、Agent的快思考-慢思考对话器-推理器架构、隐式图神经网络的图神经求解器、语言模型开发者应报告训练集-测试集重叠情况 1、[LG] Can Looped Transformers Learn to Implement Multi-step Gradient Descent for In-context Learning? K Gatmiry, N Saunshi, S J. Reddi, S Jegelka… [MIT & Google Research] 环形Transformer
………………………………
