Guess: Thinking likes by checking rotten likes taken from Traipebonacts

Discovering people's preferences is essential to building AI-ages that deliver personalized and effective interactions. Recent work has shown the possibilities of LLMs to provide user options from user interaction, but they tend to produce broad and general preferences, failing to capture a unique and individualized type. This paper introduces forecasting, a method designed to improve the accuracy and adaptability of popular trends. Forecasting includes three key elements: (1) Quantitative refinement of input preferences, (2) decomposition of preferences into selected components, and (3) validation of preferences across multiple trajectories. We test forecasting in two different environments: GridWorld editing and a new text environment (Plume). Predict more accurately the evolving human preference for improving existing infrastructure by 66.2% (gridworld environment) and 41.0% (Plume).

• † University of Colorado Boulder
• ** Work Done while at Apple