Depth Registers Unlock W4A4 on SwiGLU: A Reader/Generator Decomposition 文章

ArXiv CS.CL2026-05-26NEWSen作者: Ziyang Liu
人工智能
查看原文 →
关系图谱

Depth Registers Unlock W4A4 on SwiGLU: A Reader/Generator Decomposition · 相关技术

相关技术

ORMODEPLAlanguage model后训练zero-shotdivide-and-conquer partitioningW4A4UCTTokenSwiGLUStraight-Through EstimatorParts-of-Speech (POS) tagsNarrative Abstraction BenchmarkNPUNATLayout-as-PolicyHISGranular Alignment ParadigmGated Linear UnitsDecodeCamouflaged object detectionARGANN