Google study a flight recommendation task where an LLM acts as a flight booking assistant and interacts with a user over multiple rounds. To make good recommendations, the LLM needs to form and update its beliefs about the user's preferences.
Comments
Log in with your Bluesky account to leave a comment
They find that Bayesian Assistant model that maintains a probability distribution that reflects its beliefs about the user’s preferences and performs updates when new information becomes available, achieves better performance (a) and its predictions agree more with those of the Bayesian Assistant (b
IF I’m going to save 15 min, it’s going to be by looking at my options and picking instead of convincing and talking to an “assistant” to finally choose the flight I already decided 🤷🏻♀️
Comments
But, that’s just me.