Doc-to-LoRA and Text-to-LoRA: Instant LLM Updates

The article introduces Doc-to-LoRA and Text-to-LoRA, two complementary research papers that offer a novel approach to updating LLMs. Traditionally, updating LLMs involves either providing relevant content in the context window or fine-tuning the model. However, both of these methods are slow and expensive. Doc-to-LoRA and Text-to-LoRA address these limitations by learning an "update generator" that can be reused cheaply at deployment time.

The key innovation is the use of a hypernetwork, which is trained to generate compact LoRA adapters. At deployment time, a document or task description is fed to the hypernetwork, which returns a LoRA adapter in a single sub-second forward pass. This avoids the expensive per-task optimization pipeline entirely.

Doc-to-LoRA focuses on knowledge updates, addressing the problem of expensive context distillation. Text-to-LoRA, on the other hand, focuses on model adaptation, tackling the issue of expensive fine-tuning pipelines. Both methods share the same update cost amortization framework, with a clean separation of costs. The potential benefits of these techniques include more personalized and responsive AI systems, as well as the ability for LLMs to learn from mistakes and adapt to user preferences more effectively. The challenge will be in managing the complexity of training and ensuring the quality of the generated LoRA adapters.