Mathematical reasoning is a fundamental capability of large language models for educational use, and reliable problem solving demands not only correct answers but also coherent and verifiable reasoning steps. Multi‑agent systems exploit collaboration and cross‑checking among several LLMs, showing growing promise in this area. Existing benchmarks, however, mainly judge the final answer in text‑only settings, providing little insight into multimodal reasoning, the reliability of intermediate steps, or robustness to reasoning errors. To fill this gap we present MAMath‑Bench, a process‑aware benchmark for evaluating multimodal multi‑agent mathematical reasoning. The suite comprises 1,003 problems across 11 mathematical categories, multiple difficulty levels, and modalities such as text, images and audio. Fine‑grained milestone annotations enable assessment along three axes: correctness of the solution, reliability of the reasoning process, and robustness to adversarial perturbations of the reasoning chain. Our evaluation reveals three key findings: (1) multi‑agent collaboration does not consistently surpass strong single‑LLM accuracy; (2) collaboration markedly expands coverage of intermediate reasoning but can introduce redundant steps; (3) current systems remain vulnerable to adversarial errors, especially subtle structural perturbations. The dataset is publicly available at https://github.com/mira-ai-lab/AgentMath.
Review