MV-Physics: A Benchmark for Evaluating Multimodal Large Language Models in Physical Visual Scenes

Abstract

Vision-based multimodal foundation models exhibit strong generic reasoning capabilities, yet existing physics evaluation benchmarks fail to accommodate diagram-dependent logical constraints designed for K-12 Chinese physics education. To address this unmet testing demand, we construct MV-Physics, a layered benchmark consisting of 8,011 authentic official junior high physics exam items covering five core physical subdisciplines and three mainstream exam formats. The tiered assessment architecture of this benchmark unifies single-graph and cross-graph evaluation pipelines to deliver multi-dimensional quantification of model physical reasoning proficiency. Comparative API-based tests over mainstream multimodal models reveal Gemini 2.5 Pro attains $88.34 \%$ accuracy on isolated image tasks, while doubao-seed-1-6-vision achieves the leading 89.96% score for multi-image collaborative reasoning. Almost all evaluated models deliver subpar performance on multi-option questions, with their prediction accuracy staying below the 60% threshold. Latency statistics further indicate that most models cannot achieve an ideal tradeoff between predictive precision and inference speed during multi-graph analysis. We also extend the benchmark evaluation scope to senior high physical visual problems to verify its hierarchical discrimination capacity. As a domain-specific assessment asset, MV-Physics facilitates targeted research on multimodal physical reasoning and provides standardized competency measurement standards for compulsory education intelligent tutoring platforms.