What's this about?
ACLArena measures forgetting when capabilities are added sequentially and compares multi-teacher distillation, self-distillation and model merging.
What is catastrophic forgetting?
New training changes weights in ways that erase old skills. Replay restores prior examples, while routed LoRA experts isolate specialised updates.
Why does this matter?
Author-reported preprint; 8 HF upvotes.
- Industrial agents need multiple skills acquired at different times.
- Token-level analysis helps locate where transfer or forgetting occurs.
- Four tasks are useful evidence but not a general continual-learning solution.