Technical Report for AI4Math-2026 Track 3: VeraPhys: Verified Ensemble Repair for Multimodal Physics Reasoning

Visual physics reasoning requires systems to connect visual perception, physical laws, mathematical derivations, and faithful explanations. Although recent LLMs and VLMs have improved multimodal reasoning, they often struggle when key problem information is contained primarily in images, especially in challenging settings involving handwritten diagrams, equations, or problem statements. This technical report presents our solution for AI4Math-2026 Track 3: SeePhysPro, a multimodal physics reasoning challenge involving diagrams, expressions, and explanations. We develop VeraPhys, $\textbf{V}$erified $\textbf{E}$nsemble $\textbf{R}$ep$\textbf{A}$ir for $\textbf{Phys}$ics, a multi-stage ensemble pipeline that combines Qwen and GPT candidate generation, Python/SymPy-assisted reasoning, local-Qwen verification, OCR-enhanced visual evidence, and conservative answer repair. On the official $\texttt{test-mini}$ evaluation, our approach achieves $70.96\%$ overall accuracy.