Part 1-1

Task-wise Selective Adaptation

Complete streamIsCiL stays near the task-copy baseline at 85.2 versus 86.2 AUC, while retaining positive BWT +11.2.

Incomplete streamLocal skill removal reaches 72.7 AUC and BWT +12.4, versus 33.8 AUC for the task-copy baseline.

Overall performance after unlearning

Complete unlearning

TAIL-τ CLPU86.2AUC

FWT 86.2BWT 0.0

IsCiL (ours)85.2AUC

FWT 75.0BWT +11.2

Incomplete unlearning

TAIL-τ CLPU33.8AUC

FWT 33.8BWT 0.0

IsCiL (ours)72.7AUC

FWT 61.4BWT +12.4

The skill memory of an agent: the skills of the task being unlearned are removed as a row, while the skill retriever and the skill decoder stay frozen

Stream20 CiL stages, plus unlearning stages

  • CiL stage
  • Unlearning stage
An unlearning stage brings no demonstration to imitate.
Complete unlearning
Evolving Kitchen-Complete
Incomplete unlearning
Evolving Kitchen-Incomplete