✨
AI Summary
- Sparse memory finetuning addresses catastrophic forgetting in LLMs during continual learning by selectively training only memory slots highly activated by new knowledge using TF-IDF ranking
- Achieves new knowledge acquisition comparable to full finetuning and LoRA while substantially reducing degradation of previously acquired capabilities on held-out QA benchmarks
- Leveraging sparsity in memory layers offers promising strategy for LLMs to continually accumulate knowledge over time without forgetting prior information