Prompt optimization aims to find the best prompt to a large language model (LLM) for a given task. LLMs have been successfully used to help find and improve prompt candidates for single-step tasks. However, realistic tasks for agents are multi-step and introduce new challenges: (1) Prompt content is likely to be more extensive and complex, making it more difficult for LLMs to analyze errors, (2) the impact of an individual step is difficult to evaluate, and (3) different people may have varied preferences about task execution. While humans struggle to optimize prompts, they are good at providing feedback about LLM outputs; we therefore introduce a new LLM-driven discrete prompt optimization framework PROMST that incorporates human-designed feedback rules to automatically offer direct suggestions for improvement. We also use an extra learned heuristic model that predicts prompt performance to efficiently sample from prompt candidates. This approach significantly outperforms both human-engineered prompts and several other prompt optimization methods across 11 representative multi-step tasks (an average 10.6%-29.3% improvement to current best methods on five LLMs respectively). We believe our work can serve as a benchmark for automatic prompt optimization for LLM-driven multi-step tasks.

To collaborate effectively in physically situated tasks, robots must be able to ground concepts in natural language to the physical objects in the environment as well as their own capabilities. We describe the implementation and the demonstration of a system architecture that sup- ports tasking robots using natural language. In this architecture, natural language instructions are first handled by a dialogue management component, which provides feedback to the user and passes executable instructions along to an Abstract Meaning Representation (AMR) parser. The parse distills the action primitives and parameters of the instructed behavior in the form of a directed a-cyclic graph, passed on to the grounding component. We find AMR to be an efficient formalism for grounding the nodes of the graph using a Distributed Correspondence Graph. Thus, in our approach, the concepts of language are grounded to entities in the robot’s world model, which is populated by its sensors, thereby enabling grounded natural language communication. The demonstration of this system will allow users to issue navigation commands in natural language to direct a simulated ground robot (running the Robot Operating System) to various landmarks observed by the user within a simulated environment.