Accelerated Preference Elicitation with LLM-Based Proxies

  • 2025-01-24 16:42:47
  • David Huang, Francisco Marmolejo-Cossío, Edwin Lock, David Parkes
  • 0

Abstract

Bidders in combinatorial auctions face significant challenges when describingtheir preferences to an auctioneer. Classical work on preference elicitationfocuses on query-based techniques inspired from proper learning--often viaproxies that interface between bidders and an auction mechanism--toincrementally learn bidder preferences as needed to compute efficientallocations. Although such elicitation mechanisms enjoy theoretical queryefficiency, the amount of communication required may still be too cognitivelytaxing in practice. We propose a family of efficient LLM-based proxy designs for elicitingpreferences from bidders using natural language. Our proposed mechanismcombines LLM pipelines and DNF-proper-learning techniques to quicklyapproximate preferences when communication is limited. To validate ourapproach, we create a testing sandbox for elicitation mechanisms thatcommunicate in natural language. In our experiments, our most promising LLMproxy design reaches approximately efficient outcomes with five times fewerqueries than classical proper learning based elicitation mechanisms.

 

Quick Read (beta)

loading the full paper ...