LLM-Powered Preference Elicitation in Combinatorial Assignment

TL;DR

The paper tackles preference elicitation in high-dimensional combinatorial assignment, focusing on course allocation, by introducing an LLM-proxy framework that reduces human effort while maintaining allocation quality. It couples one-shot natural language input with a noise-robust learning pipeline, chain-of-thought prompting, and epistemic-uncertainty-driven query acquisition to guide comparison queries. The approach yields up to 20% improvement in allocative efficiency and demonstrates robustness across LLM architectures and varying user-errors, with theoretical and empirical support for using GCE loss and Double Thompson Sampling. The work highlights practical viability, cost considerations, and broad applicability to other iterative combinatorial mechanisms. This offers a scalable path to more efficient, accessible, and fair allocations in real-world settings.

Abstract

We study the potential of large language models (LLMs) as proxies for humans to simplify preference elicitation (PE) in combinatorial assignment. While traditional PE methods rely on iterative queries to capture preferences, LLMs offer a one-shot alternative with reduced human effort. We propose a framework for LLM proxies that can work in tandem with SOTA ML-powered preference elicitation schemes. Our framework handles the novel challenges introduced by LLMs, such as response variability and increased computational costs. We experimentally evaluate the efficiency of LLM proxies against human queries in the well-studied course allocation domain, and we investigate the model capabilities required for success. We find that our approach improves allocative efficiency by up to 20%, and these results are robust across different LLMs and to differences in quality and accuracy of reporting.

Figures (8)

• Figure 1: A schematic flowchart comparing Soumalias2024MLCCA with the LLM proxy algorithm in this paper.
• Figure 2: Comparison of (a) normalized allocated bundle value and (b) centered MAE, both as functions of the number of LLM-answered CQs. Shown are averages over 100 instances including 95% CIs.
• Figure 3: Comparison of the normalized allocated bundle value with and without the Chain-of-Thought (CoT) reasoning approach using LLaMA 3.1 8b model. The use of CoT results in a statistically significant improvement in the allocated bundle value.
• Figure 4: Normalized allocated bundle value as a function of the number of LLM-answered CQs. We compare GCE versus BCE as the training loss for the LLM-answered CQs. Shown are averages over 50 instances including 95% CIs.
• Figure 5: LLM CQ accuracy (evaluated using LLaMA 3.1 as the language model) as a function of the number CQs already answered by the LLM. Note that later CQs are expected to be both more informative and harder to answer, as they are generated based on an acquisition function that takes into account both the current ML model, and its epistemic uncertainty. Nonetheless, accuracy remains high even with many queries. Shown are averages over 100 instances including 95% CIs.

Theorems & Definitions (10)

• Proposition 3.1
• proof
• Remark 3.2
• Remark 3.3
• proof : \ref{['prop:true_minimizer']} Proof.
• Definition 2.1: MVNN
• Remark 2.2: Initiaization
• Definition 2.3: Epistemic MVNNs
• Remark 2.4
• Remark 2.5