2 ms·

I'm still extremely skeptical of testing LLMs on any of type of these tests. The likelihood of the model being trained on the answers to the test is too high. F

I'm still extremely skeptical of testing LLMs on any of type of these tests. The likelihood of the model being trained on the answers to the test is too high. For example, with the bar exam that people go on about - it seems to me that the internet is going to be absolutely chocked full of sample questions and answers for the bar exam. Anyone with any knowledge at all in AI knows that one of the worst things you can do is evaluate your AI using data in the training set. Of course it's going to do a good job! It's telling you what you told it! Why not just put down a SQL database and stick all the answers in it (I'm being facetious, but you get my point). We know from other research papers that people have managed to prompt the generative AI image generators to reproduce copyrighted pictures in a way that makes it very clear they were just very highly trained on that during training. It's the same with this, answers to standardized tests are always going to appear very commonly in your training data. The point of the standardized test works okay as a spot check for humans who have studied the course, but extremely badly for AI that hasn't studied the course but instead has just hoovered up every token in the course.

hgsgm 4y ago
Human students also get to see those sample questions. Is the Bar exam a bad test? Maybe. But it's how we judge the merit of new lawyers.
- jfengel 4y ago
  Humans and LLMs solve the bar exam differently. A human can't memorize the complete set of solved examples, so they have to generate it at least somewhat from scratch. It's why we give the bar exam under controlled circumstances, so that they're not just plagiarizing from existing sources. And that's what you want to test them for: the ability to apply their knowledge to novel situations. LLMs don't have memory limitations, so they can produce equivalent answers using very different means. But that means that it's a poor proxy for those novel situations. The real challenge a lawyer faces is sitting down with a client and getting a stream of verbiage, some of which is relevant and some of which isn't. In that, it's much like a programmer's job. Extracting the actual requirements is usually harder than writing the code for them. As with programming, some days are easier than others. There's a middle ground of tasks that are harder than just filling in a form, and easier than inventing a new algorithm or legal theory. Perhaps it's the "no code" of law: good enough for users to solve a lot of problems on their own, and they need to call out to a professional for more things. The bar exam could potentially be a proxy for computers to prove that they're ready to take up that middle ground. But it hasn't been tested on that yet.

hgsgm 4y ago

Human students also get to see those sample questions. Is the Bar exam a bad test? Maybe. But it's how we judge the merit of new lawyers.

jfengel 4y ago
Humans and LLMs solve the bar exam differently. A human can't memorize the complete set of solved examples, so they have to generate it at least somewhat from scratch. It's why we give the bar exam under controlled circumstances, so that they're not just plagiarizing from existing sources. And that's what you want to test them for: the ability to apply their knowledge to novel situations. LLMs don't have memory limitations, so they can produce equivalent answers using very different means. But that means that it's a poor proxy for those novel situations. The real challenge a lawyer faces is sitting down with a client and getting a stream of verbiage, some of which is relevant and some of which isn't. In that, it's much like a programmer's job. Extracting the actual requirements is usually harder than writing the code for them. As with programming, some days are easier than others. There's a middle ground of tasks that are harder than just filling in a form, and easier than inventing a new algorithm or legal theory. Perhaps it's the "no code" of law: good enough for users to solve a lot of problems on their own, and they need to call out to a professional for more things. The bar exam could potentially be a proxy for computers to prove that they're ready to take up that middle ground. But it hasn't been tested on that yet.