Multimodal AI models can describe a chart but often can't tell you what the numbers actually are.
Researchers built a benchmark using real-world charts stripped of visible data labels — the kind of charts that appear constantly in research papers and business reports. Current multimodal large language models scored reasonably well at reconstructing a table's structure, but fell apart when it came to recovering precise numerical values. That gap matters because "close enough" is not close enough when the goal is reproducibility or downstream analysis. The team then developed a training framework that treats chart reading as a progressive process, mirroring how humans scan a chart before drilling into specifics. A 7-billion-parameter model trained this way reached state-of-the-art accuracy on the benchmark.
Chart data extraction is one of those unglamorous tasks that quietly underpins a lot of scientific and business work — if you can't reconstruct the table behind a figure, you can't verify the finding or build on it. The result here suggests that raw scaling has not solved this problem, and that task-specific training design still matters more than model size alone.
For context, a 7B model beating larger general-purpose competitors is a recurring theme in AI research right now, and it keeps arriving as a quiet rebuke to the assumption that bigger always wins.