Worked on stabilizing the Mathvision evaluation workflow in the EvolvingLMMs-Lab/lmms-eval repository, focusing on improving reliability and reproducibility for model benchmarking. Addressed a key bug affecting evaluation stability, particularly for Qwen2.5VL results, by refining prompt engineering and adjusting parameter handling to reduce parsing errors and prevent unintended truncation. Leveraged Python to refactor evaluation logic, ensuring more accurate and consistent performance metrics across runs. These enhancements streamlined the evaluation process, enabling faster and more reliable model comparisons. The work emphasized robust bug fixing and model evaluation practices, supporting data-driven decision-making and facilitating future improvements in large model assessment workflows.