Accommodating human preferences is important for creating AI brokers that ship personalised and efficient interactions. Current work has proven the potential for LLMs to deduce preferences from person interactions, however they usually produce broad and generic preferences, failing to seize the distinctive and individualized nature of human preferences. This paper introduces PREDICT, a way designed to boost the precision and flexibility of inferring preferences. PREDICT incorporates three key components: (1) iterative refinement of inferred preferences, (2) decomposition of preferences into constituent elements, and (3) validation of preferences throughout a number of trajectories. We consider PREDICT on two distinct environments: a gridworld setting and a brand new text-domain atmosphere (PLUME). PREDICT extra precisely infers nuanced human preferences enhancing over present baselines by 66.2% (gridworld atmosphere) and 41.0% (PLUME).
- † College of Colorado Boulder
- ** Work carried out whereas at Apple
