Sentiment Classification on Amazon (Generalization Gap)
1.9Generalization Gap (Δ)Mask-only
Evaluation Results
| Method | Links | |
|---|---|---|
| Mask-onlyprotocol=leave-one-out zero-shot transfer, aggregation=mean over 8 seeds2026.05 | 1.9 | |
| AGMprotocol=leave-one-out zero-shot transfer, aggregation=mean over 8 seeds2026.05 | 2.1 | |
| DANNprotocol=leave-one-out zero-shot transfer, aggregation=mean over 8 seeds2026.05 | 2.5 | |
| Fishprotocol=leave-one-out zero-shot transfer, aggregation=mean over 8 seeds2026.05 | 2.7 | |
| DROprotocol=leave-one-out zero-shot transfer, aggregation=mean over 8 seeds2026.05 | 3.3 | |
| IRMprotocol=leave-one-out zero-shot transfer, aggregation=mean over 8 seeds2026.05 | 3.4 | |
| RoBERTaprotocol=leave-one-out zero-shot transfer, aggregation=averaged across all source-target pairs2026.05 | 4.5 | |
| BERTprotocol=leave-one-out zero-shot transfer, aggregation=averaged across all source-target pairs2026.05 | 5.5 |