LentEx: Generalizable Latent Entity Extraction via Synthetic Data and Instruction-Tuned LLMs

On September 7, 2026, the LentEx framework was released in the arXiv cs.CL domain. It aims to solve the problem of identifying latent entities in text, which is difficult to achieve with traditional methods. This research indicates that LentEx is the first framework to systematically handle LEE using large language models. It generates synthetic data based on templates, which are diverse and aligned with real-world distributions, thereby alleviating the shortage of annotation datasets. Experiments show that LentEx performs significantly better than state-of-the-art models on various tasks, especially achieving breakthroughs in the MTEB Clustering Benchmark and demonstrating robust generalization capabilities for unseen domains. This framework is suitable for natural language processing tasks such as Retrieval-Augmented Generation (RAG), customer profile analysis, and knowledge graph enrichment.