LLM-CER: An Interactive System for In-Context Clustering-Based Entity Resolution with Large Language Models

Abstract

Entity Resolution (ER) identifies and links records that refer to the same real-world entity. Rule-based approaches rely on explicit similarity functions, while deep learning and pre-trained language model (PLM)-based methods require large amounts of task-specific labeled data, both of which are often difficult to obtain. Large Language Models (LLMs) provide a promising alternative via zero-shot or few-shot prompting. However, existing LLM-based ER methods still adopt the pairwise matching paradigm, leading to poor scalability and high API costs on large datasets. We introduce LLM-CER (LLM-powered Clustering-based ER), an interactive, end-to-end system that instructs LLMs to perform in-context clustering of records, that is, converting ER from pairwise matching into a clustering problem , to reduce the number of costly LLM interactions while preserving high matching quality. LLM-CER contributes three elements. First, a novel in-context clustering paradigm and a systematic design-space study of factors that affect LLMs' clustering behavior (set size, within-set diversity, record variation, and ordering). Second, an extensible pipeline and UI that let users select dataset subsets, tune clustering hyperparameters, and switch LLM models before execution. Third, an interactive visualization and monitoring suite that shows hierarchical clustering steps, flag potential misclusters, log every LLM API call, and report clustering quality and efficiency metrics across models and parameter settings. In the demo, attendees will (1) run end-to-end ER on representative datasets, (2) compare metric and cost trade-offs across configurations and LLMs, and (3) inspect and correct ambiguous clusters via the visualization tools.