RI: Small: Extracting Knowledge from Language Models for Decision Making

This project aims to integrate semantic knowledge from large language models into automated decision-making and control systems while retaining reliability and robustness. The main principle of the proposed approach is to use language models to generate proposals and guidance, but still make the final decision or plan based on principled and robust planning and control methods, such that the language models are used when their semantic predictions are useful but not relied upon to always yield the correct answer. Large language models, such as ChatGPT, have garnered considerable attention in recent years due to their ability to respond to complex user queries and fulfill elaborate requests, such as writing code, composing stories, or providing educational explanations. Because of this, there is considerable interest in using them directly as decision-making systems (for example, if a language model can give “how to” instructions for repairing a car, perhaps it can also issue commands to a robot that actually repairs a car). However, there are also numerous concerns that such models might be too unreliable or too prone to generate false predictions to be useful as decision-making systems on their own. Therefore, this project aims to integrate these models into principled methods for planning and control to leverage the semantic knowledge in these models while providing a degree of robustness. This research has significant ramifications for automated decision-making systems that need to interact with complex real-world environments, where both semantic reasoning and intelligent planning are important. This includes robotic systems, including autonomous vehicles and service robots, intelligent assistants, decision support systems, and a range of automation technologies. The technical approach in this project will be based around a probabilistic formulation that ties together the ungrounded semantic predictions from language models with grounded but non-semantic predictions from learned dynamics models. In this way, probabilistic inference machinery can be used to derive algorithms that make decisions that have a high likelihood of being semantically good according to the language model and a high likelihood of being physically (dynamically) optimal according to the learned dynamics model. In practice, this principle can be instantiated in the context of both model-based and model-free reinforcement-learning systems, learned prediction systems, and planning algorithms (by formulating planning as inference). The project will explore applications of this concept to prediction, planning and control, and exploration in reinforcement learning. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.