How Much Knowledge Can You Pack into a LoRA Adapter without Harming LLM?
Sergey Pletenev, Maria Marina, Daniil Moskovskiy, Vasily Konovalov, Pavel Braslavski, Alexander Panchenko, Mikhail Salnikov
2025-02-21
Summary
This paper talks about how to add new information to large AI language models using a method called LoRA without messing up what they already know. It's like trying to teach a smart robot new facts without making it forget its old knowledge.
What's the problem?
Big AI language models are really smart, but they can only work with the information they were originally trained on. When we try to teach them new things, we risk making them forget or misuse their old knowledge. It's like trying to add new pages to a textbook without accidentally erasing or confusing the existing content.
What's the solution?
The researchers experimented with a technique called LoRA to add new facts to an AI model called Llama-3.1. They found that mixing new facts with information the AI already knows works best. However, they discovered that even this method can cause problems. The AI might start giving wrong answers to questions it used to know, or it might become overconfident and give answers even when it's not sure.
Why it matters?
This matters because as we try to make AI smarter and more up-to-date, we need to find ways to teach it new things without breaking what it already knows. Understanding these challenges helps us create better AI that can learn and grow without losing its existing capabilities. This research is crucial for developing AI that can adapt to new information while staying reliable and accurate across many different topics.
Abstract
The performance of Large Language Models (LLMs) on many tasks is greatly limited by the knowledge learned during pre-training and stored in the model's parameters. Low-rank adaptation (LoRA) is a popular and efficient training technique for updating or domain-specific adaptation of LLMs. In this study, we investigate how new facts can be incorporated into the LLM using LoRA without compromising the previously learned knowledge. We fine-tuned Llama-3.1-8B-instruct using LoRA with varying amounts of new knowledge. Our experiments have shown that the best results are obtained when the training data contains a mixture of known and new facts. However, this approach is still potentially harmful because the model's performance on external question-answering benchmarks declines after such fine-tuning. When the training data is biased towards certain entities, the model tends to regress to few overrepresented answers. In addition, we found that the model becomes more confident and refuses to provide an answer in only few cases. These findings highlight the potential pitfalls of LoRA-based LLM updates and underscore the importance of training data composition and tuning parameters to balance new knowledge integration and general model capabilities.