Modular TTT: Rethinking Test-Time Training as Composable Modules

Benchmark Model Rank Results
multiple-choice-question-answering-mcqa-on-arc-eModular TTT 1.45B (inner-product M-SiLU)#6Accuracy: 73.74
social-commonsense-reasoning-on-siqaModular TTT 1.45B (inner-product M-SiLU)#10Accuracy: 42.32
multiple-choice-question-answering-mcqa-on-arc-cModular TTT 1.45B (inner-product M)#17Accuracy: 40.78
language-modelling-on-lambadaModular TTT 1.45B (inner-product M)#78Perplexity: 12.19
question-answering-on-boolqModular TTT 1.45B (inner-product M)#85Accuracy: 62.35
question-answering-on-piqaModular TTT 1.45B (MSE M)#111Accuracy: 74.43
common-sense-reasoning-on-winograndeModular TTT 1.45B (inner-product M)#118Accuracy: 59.98
sentence-completion-on-hellaswagModular TTT 1.45B (MSE M)#156Accuracy: 58.96