They use something called neural networks to process all the words and sentences they come across, and then they spit out an answer based on what they’ve learned.

For example, let’s say you ask your LLM “What is the capital of France?” It might look at its database (which contains a ton of information) and see that the answer to this question is Paris. Then it would respond with “The capital of France is Paris!” Pretty cool, right?

But here’s where things get interesting: LLMs aren’t perfect! They can sometimes make mistakes or give you weird answers because they don’t have a full understanding of the context in which words are used. For example, if you ask your LLM “What is love?” it might respond with something like “Love is an emotion that humans experience when they feel strong affection for another person.” But what about all those other meanings of love? Like when we say “I love pizza” or “I love my cat”? Those are different kinds of love, but your LLM might not be able to distinguish between them.

So in order to make sure that our LLMs aren’t giving us false information, we need to test them and see how well they can understand language. This is where cognitive assessments come in! We ask the LLM a bunch of questions (just like you would with a human) and then we compare its answers to what we know is true or false. If it gets most of them right, then we can say that our LLM has pretty good language skills!

But here’s where things get tricky: sometimes the way we ask a question can affect how an LLM responds. For example, if you ask “What is the tallest mountain in North America?” your LLM might respond with something like “The tallest mountain in North America is Mount McKinley.” But what if you asked it “Which mountain in North America has the highest peak?” Your LLM might still give you the same answer!

This is called prompt sensitivity, and it’s a big problem for people who want to use LLMs for things like legal research or medical diagnosis. If we can’t trust our machines to give us accurate information, then what’s the point of having them in the first place?

So that’s where this paper comes in! It talks about some case studies (which are basically examples) that show how LLMs can sometimes make mistakes or misunderstand things. But it also gives us a list of DO’S and DON’TS for running cognitive assessments on these machines, which is really helpful if you want to test them out yourself!

Overall, the main takeaway from this paper is that while LLMs are pretty cool, we need to be careful when using them. They can sometimes make mistakes or misunderstand things, so it’s important to verify their answers and make sure they’re giving us accurate information. But with a little bit of care and attention, these machines have the potential to revolutionize the way we learn and understand language!