LC-ICL: Label-Guided Contrastive In-Context Learning for Robust Information Extraction

A new technique called LC-ICL aims to make large language models more reliable at extracting structured information from text by showing them not just correct examples but also carefully labeled mistakes, according to a preprint posted to the arXiv repository on June 28, 2026 [1][2]. The method, formally named Label-Guided Contrastive In-Context Learning, addresses a limitation in current few-shot approaches to information extraction, which typically provide a language model only with positive, correct demonstrations [2]. The authors argue that this neglects the instructional value of negative examples. LC-ICL constructs prompts that pair positive samples with negative samples annotated with error-cause labels, exposing the model to detailed features of why certain predictions fail and enabling it to avoid repeating those errors during inference [2]. The paper focuses on two core information-extraction tasks: named entity recognition and relation extraction [2]. The technique selects hard negative samples and the nearest positive neighbors to a given test input, then builds in-context learning demonstrations that incorporate both. The researchers report that LC-ICL outperformed previous few-shot in-context learning methods across a range of datasets, delivering what they describe as substantial performance enhancements on a broad spectrum of related tasks [2]. The preprint was posted on arXiv, an open-access repository for electronic preprints in fields including computer science, mathematics, and physics [6]. Founded in 1991, arXiv now receives approximately 24,000 submissions per month and hosts over two million articles [6]. Papers on the platform are moderated but not peer-reviewed before posting [6]. The repository also supports a framework called arXivLabs, which allows third-party developers to build experimental tools that appear on article pages, such as citation explorers and code finders [5][4]. While the LC-ICL results are presented as a step forward for robust information extraction, the preprint has not yet undergone formal peer review. The history of high-profile preprints on arXiv—such as the 2023 claims of a room-temperature superconductor known as LK-99, which were later walked back after replication attempts identified non-superconducting causes for the observed effects—underscores the preliminary nature of such unreviewed findings [8].