agentagenticagentsaianythingattentionaugmentedbertbitnetblockbytecachechaincocodecomputecontextcotdecoderdeepseekdiffusiondpoencoderenhancedevolutiongenerationgramgruguiiclinferenceinformedintelligencejourneykvlanguagelargelatentlearninglecunledgerllmllmslmmlongloralstmmambamctsmemmemorymetamobamodelmodelingmodelsmultimultimodalneednetworksnexto1openopenaioptimaloptimizationoriginalpcplanplaypredictionpreferencer1ragreasoningrepareplicationrewardrlrnnscalingscientistsearchselfslmstepsurveyswetestthinkingtimetokentokenstrainingtransformertransformersuiv2valueviavisionvlmworld