Don't Half-listen: Capturing Key-part Information in Continual Instruction Tuning

arXiv:2403.10056v1 Announce Type: new
Abstract: Instruction tuning for large language models (LLMs) can drive them to produce results consistent with human goals in specific downstream tasks. However, the process of continual instruction tuning (CIT) for LLMs may bring about the catastrophic forgetting (CF) problem, where previously learned abilities are degraded. Recent methods try to alleviate the CF problem by modifying models or replaying data, which may only remember the surface-level pattern of instructions and get confused on held-out tasks. In …

abstract arxiv catastrophic forgetting consistent continual cs.ai cs.cl drive however human information key language language models large language large language models llms part process results tasks them type