Our paper “The Non-Relational Blind Spot: Evaluating LLMs on Multi-Table Numerical Reasoning”, by Michele Luca Contalbo, Matteo Paganelli, Paolo Papotti, and Francesco Guerra, has been accepted as a Main Conference paper at EACL 2027, taking place in Athens, Greece, from March 9–14, 2027.

The paper introduces Gradino, a benchmark generation framework designed to systematically evaluate Large Language Models on numerical question answering over multiple non-relational tables. Gradino generates configurable benchmarks from scratch, enabling controlled evaluation of reasoning tasks that require combining evidence distributed across different tables without explicit relational links.

The experimental evaluation highlights the difficulties current LLMs face when performing numerical reasoning over fragmented, multi-tabular evidence, revealing important limitations in their ability to handle complex non-relational data scenarios.

The paper was accepted among 397 Main Conference papers out of 2,932 submissions, corresponding to a 13.5% acceptance rate.

The camera-ready version will be available soon. In the meantime, the paper can be accessed on OpenReview.