Decision-Oriented Dialogue for Human–AI Collaboration

“All these situations share an underlying structured decision problem in the face of uncertainty, where communicating and collaborating with others is often critical to arrive at the best solution.

Difficult decision problems like these are precisely where AI assistants could shine. (They are also a warmup for a future in which AI assistants are able to draw on trusted domain-specific AI to help make high-stakes medical, legal, financial, and business decisions.) Automated systems can handle large amounts of information and complex computations much better than humans.”

“We begin by formalizing a class of tasks, decision-oriented dialogues, in which multiple agents must communicate in order to arrive at a joint decision. They are jointly rewarded according to the quality of the decision. Each agent starts out with different information: for example, the user knows their own travel preferences, while the AI assistant has a database of flight and hotel prices. Sharing their information allows them to better assess different travel plans. Critically, however, the large amount of information and (in some tasks) the combinatorial solution space make it unnatural and inefficient for assistants to communicate all of their knowledge to users, or vice versa. Instead, agents must determine what their partners already know and what information is likely to be decision relevant, asking clarification questions and making inferences as needed.”

“Ideally, systems would be able to assist us in the comprehensive way a human travel agent would: starting with an under-specified set of “things we’d like to do,” comprehensively exploring multi-day itineraries based on the user’s preferences and domain knowledge, and iteratively refining the plan with the user based on feedback.”

“Across all task settings, current-era language models did not perform as well as humans, suggesting failures in their ability to communicate efficiently and reason in structured real-world optimization problems. Future modeling work in this domain may seek to integrate tools and inference techniques which would allow language models to compute optimal decisions for these types of problems while maintaining their flexible communication and collaboration skills.”