Charts are central to analytical reasoning, yet existing benchmarks for chart understanding focus almost exclusively on single-chart interpretation rather than comparative reasoning across multiple charts. To address this gap, we introduce ChartDiff, the first large-scale benchmark for cross-chart comparative summarization. ChartDiff consists of 8,541 chart pairs spanning diverse data sources, chart types, and visual styles, each annotated with LLM-generated and human-verified summaries describing differences in trends, fluctuations, and anomalies. Using ChartDiff, we evaluate general-purpose, chart-specialized, and pipeline-based models. Our results show that frontier general-purpose models achieve the highest GPT-based quality, while specialized and pipeline-based methods obtain higher ROUGE scores but lower human-aligned evaluation, revealing a clear mismatch between lexical overlap and actual summary quality. We further find that multi-series charts remain challenging across model families, whereas strong end-to-end models are relatively robust to differences in plotting libraries. Overall, our findings demonstrate that comparative chart reasoning remains a significant challenge for current vision-language models and position ChartDiff as a new benchmark for advancing research on multi-chart understanding.

We review the current schemes of text-image matching models and propose improvements for both training and inference. First, we empirically show limitations of two popular loss (sum and max-margin loss) widely used in training text-image embeddings and propose a trade-off: a kNN-margin loss which 1) utilizes information from hard negatives and 2) is robust to noise as all K-most hardest samples are taken into account, tolerating pseudo negatives and outliers. Second, we advocate the use of Inverted Softmax (IS) and Cross-modal Local Scaling (CSLS) during inference to mitigate the so-called hubness problem in high-dimensional embedding space, enhancing scores of all metrics by a large margin.