agentagenticagentsaianythingattentionaugmentedbertblockbytecachechaincocomputecontextcotdecoderdeepseekdiffusiondpoencoderenhancedevolutiongenerationgramgruguiiclinferenceinformedjourneykvlanguagelargelatentlearninglecunllmllmslmmlonglstmmctsmemmemorymetamobamodelmodelingmodelsmultimultimodalneednetworksnexto1openopenaioptimaloptimizationoriginalpcplanplaypredictionpreferencer1ragreasoningreplicationrewardrlrnnscalingscientistsearchselfslmstepsurveyswetestthinkingtimetokentokenstrainingtransformertransformersvalueviavisionworld