Accelerated Preference Elicitation with LLM-Based Proxies

TL;DR

This work addresses the cognitive burden of describing bidder preferences in combinatorial auctions by introducing natural-language, LLM-based proxies integrated into the Competitive Equilibrium for Combinatorial Auctions (CECA). It develops a Simulation Pipeline to model bidder behavior and evaluates multiple proxy designs, including drop-in, plus-questions, and hybrid variants, showing substantial reductions in required queries while preserving high welfare. The key finding is that LLM proxies can achieve approximately efficient outcomes with up to fivefold fewer queries than classical DNF-learning proxies, with robustness and coherence demonstrated across models and seeds. The study provides a testing sandbox and outlines future directions, including lab-human experiments, scalability, and studying strategic or adversarial behavior in natural-language elicitation.

Abstract

Bidders in combinatorial auctions face significant challenges when describing their preferences to an auctioneer. Classical work on preference elicitation focuses on query-based techniques inspired from proper learning--often via proxies that interface between bidders and an auction mechanism--to incrementally learn bidder preferences as needed to compute efficient allocations. Although such elicitation mechanisms enjoy theoretical query efficiency, the amount of communication required may still be too cognitively taxing in practice. We propose a family of efficient LLM-based proxy designs for eliciting preferences from bidders using natural language. Our proposed mechanism combines LLM pipelines and DNF-proper-learning techniques to quickly approximate preferences when communication is limited. To validate our approach, we create a testing sandbox for elicitation mechanisms that communicate in natural language. In our experiments, our most promising LLM proxy design reaches approximately efficient outcomes with five times fewer queries than classical proper learning based elicitation mechanisms.

Figures (5)

• Figure 1: Overview of our proxy auction design with people, proxies and the auctioneer. The auctioneer communicates with the proxies to run the auction, and cannot communicate with the people directly. Each person's proxy maintains a belief of its person's preferences, which it refines over time by communicating with its person using value queries, demand queries and natural language questions.
• Figure 2: Average efficiency of the auction run under the three scenarios from Table \ref{['table:scenario-items']}. Anchoring on our DNF learning proxy, we see that the drop-in LLM proxy $\omega_{\textsc{vd1}}$ performs similarly, also reaching 75% efficiency in around $10$ person--proxy interactions. The proxy $\omega_{\textsc{vd2}}$ performs significantly better, reaching 75% efficiency in four interactions; proxy $\omega_{\textsc{nvd}}$, which additionally uses a natural language question, reaches 75% efficiency in two interactions.
• Figure 3: Performance of the hybrid proxy $\omega_{\textsc{h}}$ and the DNF learning proxy $\omega_{\textsc{xor}}$ with respect to number of person--proxy interactions. Panel (a) shows that $\omega_{\textsc{h}}$ achieves a good approximation of its person's valuation significantly quicker than $\omega_{\textsc{xor}}$, while still converging to the person's exact valuation. Panel (b) shows a rapid efficiency increase initially for $\omega_{\textsc{h}}$, and similar long-run efficiency.
• Figure 4: Comparison between GPT-based and other LLM models models in person robustness tests. Gpt-4o-mini and other LLM-models give similar values in response to value queries evidenced by the dots, each of which representing a single bundle, falling near the identity gpt-4o-mini value ($) = Other LLM value line ($).
• Figure 5: Preference variability and shape after iteratively adding items to an initially empty bundle, and then iteratively removing items. Each box plot displays the distribution of values obtained from with 10 iterations of the process.