| common-sense-reasoning-on-winogrande | Maglev (lambda=0.1) | – | Accuracy: 58.7 |
| language-modelling-on-lambada | Maglev (separate parameters, lambda=0.1) | – | Accuracy: 48.7Perplexity: 7.73 |
| multiple-choice-question-answering-mcqa-on-arc-e | Maglev (lambda=0.1) | – | Accuracy: 70.0 |
| question-answering-on-boolq | Maglev (separate parameters, lambda=1) | – | Accuracy: 64.0 |
| question-answering-on-piqa | Maglev (separate parameters, lambda=1) | – | Accuracy: 72.6 |
| social-commonsense-reasoning-on-siqa | Maglev (separate parameters, lambda=0.1) | – | Accuracy: 42.7 |