The score results of Qwen-VL on OmniDocBench are inconsistent.

In Qwen's technical report, Qwen2.5-VL-7B scores 0.308, while the 72B version scores 0.226. [report pdf](https://www.52nlp.cn/wp-content/uploads/2025/02/Qwen2.5-VL%E6%8A%80%E6%9C%AF%E6%8A%A5%E5%91%8A.pdf)

However, in the DeepSeek-OCR paper, the 7B model achieves 0.316, whereas the 72B model achieves 0.214.

Is this a reasonable range of variation, or were any post-processing parameters modified when you re-ran the inference? Waiting for your reply, Thank you very much!