MAP
MAP coordinates LLM-powered agents to personalize for several people at once. Agents gather each person's rules, a Planner plans around the conflicts between them and explains its choices, and the user checks the plan and gives feedback. In the paper's evaluation, MAP retrieved rules and resolved conflicts more accurately than a single LLM. Try it on the assistive care scenario from the paper.
The agents' responses are scripted for this demo, including one rule the agents fail to retrieve, a kind of failure participants found in the study. The residents' documents and the Planner's instruction for resolving conflicts are the ones the system used.
The task: assistive care
A care facility's robot looks after three residents, Blaine, Ryan, and Susie. Each resident's document lists 20 preferences and schedule items. Their needs often overlap, and when they do, the robot's instructions say to serve residents in alphabetical order of their first names.
Ask the robot how it will care for them, and check its plan.
Residents' documents
Planner context
Instruction in its system prompt. When residents' preferences or schedules overlap, serve them in alphabetical order of their first names.