agentagenticagentsaianythingattentionaugmentedbertblockbytecachechaincocomputecontextcotdecoderdeepseekdiffusiondpoencoderenhancedevolutiongenerationgramgruguiiclinferenceinformedjourneykvlanguagelargelatentlearninglecunllmllmslmmlonglstmmctsmemmemorymetamobamodelmodelingmodelsmultimultimodalneednetworksnexto1openopenaioptimaloptimizationoriginalpcplanplaypredictionpreferencer1ragreasoningrepareplicationrewardrlrnnscalingscientistsearchselfslmstepsurveyswetestthinkingtimetokentokenstrainingtransformertransformersvalueviavisionworld