Speaker
Florian Mai
Description
Large language models are strong heuristic reasoners, but their planning abilities remain poor. We introduce a method for language models to learn to plan from unlabeled data by using a planner model to predict many steps ahead and conditioning the language model on the predicted plans. A crucial parameter in this framework is the level of abstraction of the generated plans: While some tasks arguably benefit more from high-level planning (e.g. creative writing), others require planning in the concrete language space (e.g. mathematical reasoning). In this talk, we explore both ends of the spectrum and finally ask the question if the right granularity can be learned from data.
