Language Model Pre-training Improves Generalization in Policy Learning

Shuang Li, Xavier Puig, Yilun Du, Ekin Akyürek, Antonio Torralba, Jacob Andreas, Igor Mordatch

Keywords: Large Language model, Imitation learning, Interactive tasks, policy learning

Abstract: Language model (LM) pre-training has proven useful for a wide variety of language processing tasks, including tasks that require nontrivial planning and reasoning capabilities. Can these capabilities be leveraged for more general machine learning problems? We investigate the effectiveness of LM pretraining to scaffold learning and generalization in autonomous decision-making. We use a pre-trained GPT-2 LM to initialize an interactive policy, which we fine-tune via imitation learning to perform interactive tasks in a simulated household environment featuring partial observability, large action spaces, and long time horizons. To leverage pre-training, we first encode observations, goals, and history information as templated English strings, and train the policy to predict the next action. We find that this form of pre-training enables generalization in policy learning: for test tasks involving novel goals or environment states, initializing policies with language models improves task completion rates by nearly 20%. Additional experiments explore the role of language-based encodings in these results; we find that it is possible to train a simple adapter layer that maps from observations and action histories to LM embeddings, and thus that language modeling provides an effective initializer even for tasks with no language as input or output. Together, these results suggest that language modeling induces representations that are useful for modeling not just language, but natural goals and plans; these representations can aid learning and generalization even outside of language processing.

One-sentence Summary: We investigate the effectiveness of LM pretraining to scaffold learning and generalization in autonomous decision-making.