Evaluating Numeracy of Language Models as a Natural Language Inference Task

Rahmad Mahendra, Damiano Spina, Lawrence Cavedon, Karin Verspoor

Abstract: While recent advancements in large language models (LLMs) have enhanced their capabilities to solve mathematical problems, other aspects of numeracy remain underexplored. In this paper, we propose a benchmark to evaluate the ability of language models to perform basic numeracy tasks. We frame numeracy as a Natural Language Inference (NLI) task to assess the models’ ability to understand both numbers and language contexts. We evaluate 49 language models (LMs), including fine-tuned LMs on NLI datasets, instruction-tuned LLMs, and specialized math-LLMs. Our findings reveal three main insights: (1) LLMs only clearly outperform smaller LMs in arithmetic tasks, indicating that mathematical reasoning cannot be generalized to other numeracy skills such as number comparison and normalization; (2) while most language models achieve fair to good accuracy for NLI entailment cases, they still struggle to predict contradiction and neutral cases; and (3) the robustness of language models’ numeracy capabilities needs improvement, particularly in understanding the semantics and pragmatics of numbers in linguistic contexts.