ChartCF achieves strong chart understanding performance in VLMs using significantly less training data by generating code-based counterfactuals, selecting similar samples, and performing multimodal preference optimization.
ECD fails to correctly identify the subplot location and misreads the axis range, while ChartCF correctly locates the target subplot and identifies the highest tick value as 1.5
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Learning More from Less: Exploiting Counterfactuals for Data-Efficient Chart Understanding
ChartCF achieves strong chart understanding performance in VLMs using significantly less training data by generating code-based counterfactuals, selecting similar samples, and performing multimodal preference optimization.