Natural Language Understanding on GLUE (CoLA, MRPC, QNLI, MNLI, RTE Subset)
66.62CoLA ScoreSTAR
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| STARK=8, k=4, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 66.62 | 89.68 | 92.61 | 86.74 | 75.57 | 82.24 | |
| No Interpolation (STAR ablation)K=8, k=4, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 66.02 | 89.56 | 92.39 | 86.51 | 73.53 | 81.6 | |
| Random basis (STAR ablation)K=8, k=4, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 65.8 | 90.2 | 92.54 | 75.52 | 73.89 | 79.59 | |
| STARK=16, k=4, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 65.74 | 90.2 | 92.61 | 86.59 | 75.41 | 82.11 | |
| STARK=8, k=1, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 65.54 | 90.25 | 92.56 | 86.52 | 74.61 | 81.9 | |
| DynMoEK=9, k=7.1, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 65.17 | 90.64 | 92.59 | 86.37 | 73.41 | 81.64 | |
| Cosine RouterK=8, k=4, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 64.94 | 89.74 | 92.52 | 86.57 | 75.09 | 81.77 | |
| STARK=16, k=1, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 64.69 | 90.03 | 92.57 | 86.64 | 73.65 | 81.52 | |
| No R (STAR ablation)K=8, k=4, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 64.57 | 90.12 | 92.54 | 86.38 | 72.56 | 81.23 | |
| Cosine RouterK=16, k=4, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 64.12 | 89.74 | 92.65 | 86.59 | 75.33 | 81.69 | |
| Cosine RouterK=8, k=1, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 64.1 | 90.14 | 92.48 | 86.56 | 73.04 | 81.26 | |
| Cosine RouterK=16, k=1, Backbone=BERT-large, Evaluation Protocol=Fine-tuning2026.06 | 63.63 | 89.81 | 92.39 | 86.63 | 74.01 | 81.29 |