Natural Language Understanding on GLUE BERT-base experts (test) (RTE, MRPC, QNLI, CoLA Subset)
60.7RTE ScoreExperts (UB)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Experts (UB)Backbone=BERT-base, Protocol=Upper Bound (single-task experts)2025.12 | 60.7 | 88.3 | 81.6 | 56.2 | 71.7 | |
| UMTAMBackbone=BERT-base2025.12 | 57.8 | 85.5 | 78.2 | 48.7 | 67.5 | |
| TIESBackbone=BERT-base, Sparsity (k)=20%2025.12 | 57.4 | 85.1 | 78.1 | 45.2 | 66.4 | |
| LinearBackbone=BERT-base, Merging=Task Arithmetic2025.12 | 56 | 80.7 | 64.5 | 33.9 | 58.8 | |
| DARE+TIESBackbone=BERT-base, Sparsity (k)=20%2025.12 | 56 | 85.9 | 78.5 | 35.3 | 63.9 | |
| DAREBackbone=BERT-base, Sparsity (k)=20%2025.12 | 54.9 | 79.5 | 63.8 | 44.3 | 60.6 |