Q&A: Tom Hartvigsen on Making Large Language Models More Reliable
Tom Hartvigsen is among the new class of faculty who joined the University of Virginia’s School of Data Science this fall. He spoke recently about what brought him to UVA as well as a new method he published to keep large language models, like ChatGPT, up to date, which will soon be featured at NeurIPS, one of the most prestigious publication venues for machine learning.
Q. Tell me about what drew you to UVA’s School of Data Science.
I think the main thing is the opportunity to target impact. Coming from a more traditional computer science setting, it can be hard to do work targeting impact in health care, for instance. But in reality, impactful work in health care or any other domain should start with understanding their problems, then building the solutions that they really need.
And I think that it’s hard to reward that in more traditional departments, where it’s easier to count publications instead of evaluating more holistically based on impact-focused research.
Q. You recently had a paper accepted by a top machine learning publication venue, which you’ll be presenting later this year. In it, you propose a new method for keeping large language models up to date. Can you describe why this is so important?
The key motivation is that large language models have become so big and expensive, that when you identify things that you don’t like, like misbehaviors or misinformation, it’s really hard to fix them in a cost-effective way. And it’s also a big power imbalance because very few people can actually afford to train and retrain these models. So, this is also a direction for decentralizing the control of what knowledge is in different models because not everybody can even afford to do the retraining – but you can afford to do small updates.
Q. How does your method work?
The idea is that you’ve deployed your model and you’ve identified some mistake – it used to predict that and now it should predict this. We propose a method called GRACE, which introduces a new module you can plug inside your big model, and every time you pass new information into the model, the adapter can store it as a memory. So, when you want to update your model, you just have to add the mistake to the adapter’s memory along with the new fix. When new inputs resemble previous mistakes, the adapter can then apply a fix, letting you efficiently keep your model up to date.
Q. For folks who may not know much about large language models, what do you think they should know about their impact?
A large language model is a machine learning model that you can pass text into, and it’ll give you back text. And nowadays, they are extremely good at generating text, so people can’t tell it’s from a machine. One place that this has been impactful so far is in writing code. You can basically type in “write me a function that does this thing,” and it’ll give you back correctly formatted code that will actually run and do the stuff that you want it to.
You can prompt these models and ask them questions, and they will return very human, readable text because they’ve already seen most human-written text. But just because they are so big and they perform well at a huge range of tasks doesn’t mean that they’re not still machine learning models with the same problems as every other machine learning model. They still encode the social norms present in the training data. And similarly, they’re trained on biased language, and so they very easily generate biased language against minority groups.
Q. What are some of your other research priorities in the next year?
Beyond responsible language modeling, a big portion of my research has always been on time series data, where we’re building methods for learning from data collected over time. Using these data, we try to predict what’s going to happen in the future or other sorts of tasks. And so one of my group’s main efforts is on learning generalizable knowledge about the world through multiple sources of time series data.
Doing this is important for a million reasons, but one reason is that language models, and pretty much all of machine learning, sort of operates in this “flat earth” view of the world, where our data don’t change over time. But in reality, our data are changing constantly, but there are not a lot of effective ways to incorporate change over time into our models. One great way is to explicitly focus on training models that can operate in a dynamically changing environment, especially through time series data.
Q. Can you tell us something about yourself that might surprise people?
I’ve never eaten a hamburger or hot dog. I was raised vegetarian and then went vegan.
