Math Takes Two: Benchmark Tests Emergent Mathematical Reasoning in Language Models

A novel benchmark named Math Takes Two has been introduced to evaluate the ability of language models to cultivate mathematical reasoning from foundational principles via communication. In contrast to current assessments that depend on symbolic problems rooted in established mathematical norms, this benchmark examines two agents lacking prior mathematical knowledge as they create a shared symbolic protocol to tackle a visually grounded task, where the use of numerical systems aids in extrapolation. This initiative is driven by the theory that human mathematical thinking evolved alongside the necessity for precise communication. The benchmark seeks to differentiate genuine mathematical reasoning from mere statistical pattern recognition based on formal syntax. The paper can be found on arXiv with the identifier 2604.21935.