"Planner model generates a task-specific inference program that is executed by a population of Follower models. Our approach equips LMs with the ability to write recursive search procedures that guide LM inference, enabling new forms of verifiable and efficient reasoning."
I'm a fan of several of these authors. Curious how much progress can be made with methods like this, despite the limited reasoning of the component models.
On the other hand, this might much how humans work. My reasoning is pretty limited without tools and process to ground me. And these things have usually been developed by others.
"Planner model generates a task-specific inference program that is executed by a population of Follower models. Our approach equips LMs with the ability to write recursive search procedures that guide LM inference, enabling new forms of verifiable and efficient reasoning."
I'm a fan of several of these authors. Curious how much progress can be made with methods like this, despite the limited reasoning of the component models.
On the other hand, this might much how humans work. My reasoning is pretty limited without tools and process to ground me. And these things have usually been developed by others.