OpenGVLab lists InternVL3.5 as a multi-modal model series. This article focuses on the visual question answering application of 8B checkpoint. The official research results and ChinaOPC’s actual business tests are different pieces of evidence, and this article does not report self-test results.
Read the chart structure first and then answer the trend
To simulate a monthly sales chart, you first need to identify the horizontal and vertical axis units, legend and time range, and then describe the changes. If the image has units cropped off, it should indicate a lack of information rather than an empirical inference. Dual vertical axes and stacked plots should be added to the test set separately.
Use checkable material for key figures
The numerical reading in the image can be used as a candidate, and the original table data is given priority when calculations are involved. Include chart page numbers or area descriptions in your answers so reviewers know where the judgment comes from. Seeing correlated changes cannot be directly interpreted as cause and effect.
Evaluation cannot only select clear screenshots
Testing should cover scaling, compression, dense labeling, and color proximity, recording error types and manual correction times. If the data contains customer business information, the processing authority should be confirmed first. Suitability for deployment depends on specific tasks, equipment and review arrangements, rather than a single public list.
Extended reading: AI model.