ChartSynth: Structured Summarization and Comparison of Multi-Chart Scientific Figures

Neelu Verma

Abstract: Multi-panel chart figures are ubiquitous in scientific publications, where multiple coordinated visualizations are used to compare experimental conditions, characterize parameter-dependent behavior, and expose complementary trends. However, existing chart understanding benchmarks and vision-language models predominantly operate on individual charts, limiting their ability to reason over relationships distributed across multiple panels. We introduce multi-chart summarization, a new task that requires jointly interpreting a set of charts and generating a concise summary of their similarities, differences, and cross-panel relationships. To study this task, we present MultiChartSum, a benchmark containing over 10,000 multi-panel scientific figures and 30,000+ chart panels paired with comparative summaries. We further propose ChartSynth, a modular vision-language framework that independently encodes chart panels, performs cross-panel visual-semantic alignment, and generates contrastive summaries through structured representation fusion. Unlike panel-wise captioning approaches, ChartSynth explicitly models inter-panel dependencies to distinguish shared trends from condition-specific changes. On the MultiChartSum test set, ChartSynth achieves 41.2 ROUGE-L, 89.5 BERTScore, and 75.6 Trend-F1, improving over the strongest LLaVA-MultiInput baseline by 5.6, 2.1, and 6.1 points, respectively. Human evaluation further shows that ChartSynth achieves 89.5% fluency, 79.6% insightfulness, and 76.8% comparative accuracy, substantially outperforming the evaluated baselines. These results demonstrate that explicit cross-chart alignment and relational reasoning are critical for generating scientifically meaningful summaries of multi-panel visualizations. MultiChartSum and ChartSynth establish a benchmark and modeling framework for structured multi-chart understanding and comparative vision-language reasoning.

Keywords: Multi-chart summarization, Chart understanding, Vision-language models, Cross-chart reasoning, Scientific visualization, Multimodal learning, Comparative summariza-tion, ChartSynth

How to Cite?: Neelu Verma, "ChartSynth: Structured Summarization and Comparison of Multi-Chart Scientific Figures", Volume 15 Issue 9, September 2026, International Journal of Science and Research (IJSR), Pages: 1230-1238, https://www.ijsr.net/getabstract.php?paperid=SR26919202852, DOI: https://dx.doi.org/10.21275/SR26919202852

Download Citation: APA | MLA | BibTeX | EndNote | RefMan