a1agentagenticagentsaianythingattentionaugmentedbertbitnetblockbytecachechaincocodecomputecontextcotdecoderdeepseekdefidiffusiondltdpoencoderenhancedevolutiongenerationgptgramgruguiiclinferenceinformedintelligencejourneykvlanguagelargelatentlearninglecunledgerllmllmslmmlongloralstmmambamctsmemmemorymetamimblewimblemobamodelmodelingmodelsmultimultimodalneednetworksnextnfto1openopenaioppooptimaloptimizationoriginalpcplanplaypredictionpreferencer1ragreasoningrepareplicationrewardrlrnnscalingscientistsearchselfslmstepsurveyswetestthinkingtimetokentokenstrainingtransformertransformersuiv2valueviavisionvlmworld