Lifelong Learning on SST, QA-SRL, and WOZ Permuted Sequences GPT-2 models (test)
82.4Accuracy (SRL WOZ SST)LSEBMCL
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| LSEBMCLSampling Ratio (GEN)=0.2, Evaluation Protocol=last epoch of the last task2025.12 | 82.4 | 83.1 | 82.5 | 82.7 | 83.7 | 83.2 | 82.9 | 0.5 | |
| PPSEBMSampling Ratio (GEN)=0.2, Evaluation Protocol=last epoch of the last task2025.12 | 82.2 | 83.4 | 84 | 82.2 | 85.5 | 83.3 | 83.4 | 1.2 | |
| RVAE-LAMOLSampling Ratio (GEN)=0.2, Evaluation Protocol=last epoch of the last task2025.12 | 82 | 81.5 | 82.3 | 81.5 | 81.1 | 82.8 | 81.9 | 0.6 | |
| LSEBMCLSampling Ratio (GEN)=0.05, Evaluation Protocol=last epoch of the last task2025.12 | 81.9 | 82.8 | 81.8 | 84.4 | 82.7 | 80.5 | 82.4 | 1.3 | |
| Multitaskeddescription=Simultaneous training on all tasks2019.09 | 81.5 | — | — | — | — | — | — | — | |
| LAMOL REALgamma (γ)=0.2, replay type=real samples2019.09 | 81.2 | 81.8 | 80.6 | 81.6 | 80.4 | 80.5 | 81 | 0.5 | |
| LAMOL REALgamma (γ)=0.05, replay type=real samples2019.09 | 80.9 | 81 | 78.9 | 80.1 | 77.7 | 78 | 79.4 | 1.2 | |
| PPSEBMSampling Ratio (GEN)=0.05, Evaluation Protocol=last epoch of the last task2025.12 | 80.9 | 82.8 | 83.7 | 82.1 | 85.2 | 83 | 83 | 1.4 | |
| LAMOLSampling Ratio (GEN)=0.2, Evaluation Protocol=last epoch of the last task2025.12 | 80.3 | 80.8 | 81.1 | 81.2 | 78.9 | 81.9 | 80.7 | 1 | |
| RVAE-LAMOLSampling Ratio (GEN)=0.05, Evaluation Protocol=last epoch of the last task2025.12 | 80.2 | 80.7 | 79.6 | 79.9 | 79.6 | 78 | 79.7 | 0.9 | |
| LAMOL GENgamma (γ)=0.2, replay type=pseudo-samples2019.09 | 78.7 | 80 | 80.7 | 79.6 | 78.4 | 80.5 | 79.7 | 0.8 | |
| LAMOL TASKgamma (γ)=0.2, replay type=task-specific pseudo-samples2019.09 | 78.7 | 79.4 | 79.9 | 80.1 | 79.8 | 79 | 79.5 | 0.5 | |
| LAMOL TASKgamma (γ)=0.05, replay type=task-specific pseudo-samples2019.09 | 74.7 | 77.3 | 76.9 | 78.1 | 73.4 | 75.8 | 76 | 1.5 | |
| LAMOL GENgamma (γ)=0.05, replay type=pseudo-samples2019.09 | 73.7 | 79.6 | 78.9 | 73.1 | 68.6 | 75.7 | 74.9 | 3.4 | |
| LAMOLSampling Ratio (GEN)=0.05, Evaluation Protocol=last epoch of the last task2025.12 | 73.7 | 81.5 | 79.5 | 74.8 | 70.2 | 75.8 | 75.9 | 4.1 | |
| LSEBMCLSampling Ratio (GEN)=0, Evaluation Protocol=last epoch of the last task2025.12 | 66.8 | 66.4 | 57.7 | 77.7 | 64.6 | 65.6 | 66.5 | 6.4 | |
| PPSEBMSampling Ratio (GEN)=0, Evaluation Protocol=last epoch of the last task2025.12 | 62.8 | 78.7 | 54.6 | 75.6 | 61.4 | 61.6 | 65.8 | 9.3 | |
| LAMOL TASKgamma (γ)=0, replay type=task-specific pseudo-samples2019.09 | 41.9 | 41 | 33.5 | 50.1 | 49.3 | 41.5 | 42.9 | 5.2 | |
| LAMOLSampling Ratio (GEN)=0, Evaluation Protocol=last epoch of the last task2025.12 | 39.4 | 47.1 | 38.4 | 57.8 | 45.5 | 46.6 | 45.8 | 7 | |
| LAMOL GENgamma (γ)=0, replay type=pseudo-samples2019.09 | 38.6 | 46.5 | 36.6 | 56.6 | 44.9 | 45.2 | 44.8 | 6 | |
| EWCEvaluation Protocol=last epoch of the last task2025.12 | 37.5 | 51 | 50 | 66.4 | 44.8 | 40.7 | 48.4 | 10.2 | |
| EWCmode=Online2019.09 | 35.5 | 50.6 | 48.4 | 64.7 | 43.9 | 39 | 47 | 8.7 | |
| Fine-tunedEvaluation Protocol=last epoch of the last task2025.12 | 33.2 | 52 | 25.5 | 64.6 | 34.6 | 34.6 | 40.8 | 14.6 | |
| GEMEvaluation Protocol=last epoch of the last task2025.12 | 32.9 | 52.1 | 31.5 | 64 | 45.3 | 36.7 | 43.8 | 12.6 | |
| GEMmemory size=5% of current task size2019.09 | 32.6 | 50.4 | 29.8 | 63.3 | 44.1 | 36.3 | 42.8 | 11 | |
| MASEvaluation Protocol=last epoch of the last task2025.12 | 32.3 | 37.3 | 46.2 | 57.3 | 50.3 | 32.2 | 42.6 | 10.3 | |
| Fine-tuneddescription=Direct fine-tuning on the stream of tasks2019.09 | 31.3 | 50.2 | 24.7 | 62.9 | 32.8 | 33.9 | 39.3 | 12 | |
| MAS2019.09 | 31 | 36.5 | 45.3 | 56.6 | 49.7 | 30.8 | 41.6 | 8.9 | |
| MultitaskedRole=Upper bound2025.12 | — | — | — | — | — | — | 87 | — |