MathModelingAgent: An Evidence-Driven Verification Plugin for Mathematical Modeling and Competition Problems

MathModelingAgent is a DSH plugin designed for open-ended mathematical modeling, prediction, optimization, and other tasks, addressing the issue where code runs correctly but conclusions may still be erroneous. Its core mechanism registers key conclusions as Claims, which are then verified through Obligations and Evidence, explicitly distinguishing among three states: PASS, FAIL, and INCONCLUSIVE, with particular emphasis on retaining the INCONCLUSIVE state when evidence is insufficient to avoid forcing a definitive conclusion. The plugin provides two Skills: math-modeling-agent for advancing the model, and math-modeling-audit for independent auditing without modifying the original text. It supports optional tools such as Python, Lean, and Wolfram for numerical computation.