Math Takes Two: A test for emergent mathematical reasoning in communication

Research paper proposes "Math Takes Two," a novel benchmark that tests emergent mathematical reasoning in language models by requiring two agents without prior mathematical knowledge to develop a shared symbolic protocol to solve visually grounded tasks. Unlike conventional benchmarks relying on established mathematical conventions, it evaluates whether models discover abstract concepts from scratch. The work directly probes whether LLMs demonstrate true mathematical reasoning or merely pattern matching over formal syntax.