Gretel Navigator Research - Agent-Based Programming Unlocks New Possibilities for LLMs

Speaker 1 (00:00):

This is Alex from Gretel. Today I'm going to walk through some really exciting research that we are doing for the TAP LM model. So I'm going to show an example where we're going to ask an LM here to solve a physics problem at the high school or college level. Essentially, it's using kinematic equations. We are generating a data frame, which has position and time, and then we're going to ask the LM to see, given that position and time, if it can, calculate the velocity and the acceleration at each point. So here you see the data. F we're creating time and position. We're giving it a prompt. We're saying, here's the velocity and definition of velocity. Here's the acceleration definition of acceleration. We're passing it off to our tab LM application. So we're giving it an input and the prompt, and it's already done running. So it's pretty fast.

(00:46)
And let's look at the answers here. So interesting. So the LM did a fairly good job at estimating what the results are going to be. Now we get into something that is just super exciting. So what we're going to do is we are going to take advantage of something called an agent or plan and execution based approach where we can ask the LM to write code to solve this math problem instead of just guessing as LMS do about what the correct values are. So now essentially the LM is going to be writing code, executing that code, and programmatically calculating the results here. So this is a pretty cool challenge. Let's do head sell and we'll run all below. So we'll rerun these so we can see our prompt going through these calc tags. Tell our LLM tell to type LM application essentially to treat this as a formula that is running. So here you see it's taking a little bit longer to run. That's because it's running essentially this plan and execution based approach. So right now it should be thinking and writing code, and it wrote two different columns programmatically. Let's look at the results. Nailed it. Perfect results. So really exciting coming soon with the tab LM.