John Schulman
TalkRL: The Reinforcement Learning Podcast
00:00
Do You Think the RL Task Is Going to Work?
We had two phases in each episode of the RL task. So there was first a browsing phase where the model goes and issues searches and clicks on things, quotes relevant information. And then once it's done browsing, it'll issue another command called end browsing and it will write its answer. That's also expressed in tokens. But really, we rolled this all into one big RL task where your episode involves browsing and writing out the answer. It's all one big RL episode. Did you think this is going to work well or were you kind of surprised? At the very beginning of the project, we didn't know if it was going to work or not.
Transcript
Play full episode
