The iDEA-iSAIL Joint Laboratory at the University of Illinois received an Outstanding Paper Award at the International Conference on Learning Representations (ICLR) 2026 Logical Reasoning of Large Language Models Workshop for their paper, "RAG Over Tables: Hierarchical Memory Index, Multi-State Retrieval, and Benchmarking." Paper authors include lab members Jingrui He, professor and MSIM program director; Sirui Chen, Xinrui He, and Zihao Li, computer science PhD students; Jiaru Zou, computer science MS student; Dongqi Fu, alum; as well as Jiawei Han, professor of computer science, and Yada Zhu, IBM collaborator. Chen gave an oral presentation of the research at the workshop, which was held last month in Rio de Janeiro, Brazil. This award was selected out of 206 accepted papers at the workshop.
With a large amount of real-world knowledge stored in tables across websites, databases, and documents, current AI systems often struggle to locate and synthesize relevant information when a user poses a question whose answer is distributed across multiple tables. The lab's latest work, T-RAG, addresses this challenge by organizing large-scale table collections into a structured graph that captures relationships between tables from multiple perspectives, including their meaning, their format, and their vocabulary.
"When a query arrives, T-RAG applies a multi-stage retrieval process to efficiently identify the most relevant tables, then provides them along with their relational context to a large language model for reasoning and answer generation," explained Chen. "In our paper, we also introduce MultiTableQA, the first benchmark designed to evaluate cross-table question answering using real-world table sources and human-annotated queries."