Sequential adaptationSeq-FT and Seq-LoRA forget what they learn (BWT −63.7 / −55.2), and EWC avoids it only by learning less.
Adapter-basedTask-level adapters are already competitive: IsCiL leads Kitchen at 89.8 AUC, while TAIL-τ leads World at 85.7 versus 84.3.
| CiL algorithm | Evolving Kitchen | Evolving World |
|---|
| FWT (%) | BWT (%) | AUC (%) | FWT (%) | BWT (%) | AUC (%) |
|---|
| Pre-trained | – | – | 24.3 | – | – | 0.0 |
|---|
| Sequential adaptation | Seq-FT | 90.9 | −63.7 | 35.0 | 88.9 | −73.6 | 24.9 |
|---|
| EWC | 34.2 | −19.5 | 17.1 | 25.7 | −18.0 | 10.5 |
|---|
| Seq-LoRA | 77.5 | −55.2 | 28.3 | 85.6 | −75.1 | 21.4 |
|---|
| Adapter-based | L2M | 24.7 | −2.5 | 22.7 | 72.1 | −6.6 | 65.9 |
|---|
| L2M-g | 38.2 | −6.5 | 32.3 | 64.2 | −19.3 | 48.6 |
|---|
| TAIL-g | 85.3 | −49.9 | 41.5 | 90.0 | −56.8 | 39.5 |
|---|
| TAIL-τ | 86.2 | 0.0 | 86.2 | 85.7 | 0.0 | 85.7 |
|---|
| IsCiL (ours) | 79.3 | 11.0 | 89.8 | 81.7 | 2.7 | 84.3 |
|---|
| Oracle | Multi-task | 93.3 | −1.6 | 92.3 | 88.6 | 2.8 | 90.7 |
|---|
FWTnew tasks of the stageBWTtasks already learnedAUCoverall, across 20 stages