Part 1-1

Task-wise Selective Adaptation

Adapter-basedA task-level adapter no longer fits a partial demonstration: TAIL-τ reaches 41.2 AUC, while TAIL-g falls to BWT −21.1.

Skill memoryIsCiL updates only the skills a demonstration covers: BWT +8.6 / +9.3 at 75.8 / 68.9 AUC.

CiL algorithm		FWT (%)	BWT (%)	AUC (%)	FWT (%)	BWT (%)	AUC (%)
CiL algorithm		Evolving Kitchen			Evolving World
	Pre-trained	–	–	29.1	–	–	0.0
Sequential adaptation	Seq-FT	37.1	−25.1	16.5	38.9	−27.5	13.2
	EWC	27.2	−18.0	12.2	13.9	−9.1	6.2
	Seq-LoRA	37.4	−25.5	15.9	32.2	−18.2	16.0
Adapter-based	L2M	19.2	0.2	19.1	41.0	6.3	47.0
	L2M-g	37.9	−4.5	32.1	44.5	3.4	48.2
	TAIL-g	55.0	−21.1	37.2	43.2	−17.6	27.4
	TAIL-τ	41.2	0.0	41.2	27.5	0.0	27.5
	IsCiL (ours)	68.1	8.6	75.8	60.0	9.3	68.9
Oracle	Multi-task	75.4	8.0	83.2	55.0	27.6	80.9

FWTnew tasks of the stageBWTtasks already learnedAUCoverall, across 20 stages