INTERCHART: Benchmarking Visual Reasoning Across Decomposed and Distributed Chart Information

---

🧩 Overview

INTERCHART is a multi-tier benchmark that evaluates how well vision-language models (VLMs) reason across multiple related charts, a crucial skill for real-world applications like scientific reports, financial analyses, and policy dashboards.
Unlike single-chart benchmarks, INTERCHART challenges models to integrate information across decomposed, synthetic, and real-world chart contexts.

Paper: INTERCHART: Benchmarking Visual Reasoning Across Decomposed and Distributed Chart Information

---

📂 Dataset Structure

INTERCHART/
├── DECAF
│   ├── combined       # Multi-chart combined images (stitched)
│   ├── original       # Original compound charts
│   ├── questions      # QA pairs for decomposed single-variable charts
│   └── simple         # Simplified decomposed charts
├── SPECTRA
│   ├── combined       # Synthetic chart pairs (shared axes)
│   ├── questions      # QA pairs for correlated and independent reasoning
│   └── simple         # Individual charts rendered from synthetic tables
├── STORM
│   ├── combined       # Real-world chart pairs (stitched)
│   ├── images         # Original Our World in Data charts
│   ├── meta-data      # Extracted metadata and semantic pairings
│   ├── questions      # QA pairs for temporal, cross-domain reasoning
│   └── tables         # Structured table representations (optional)

Each subset targets a different level of reasoning complexity and visual diversity.

---

🧠 Subset Descriptions

1️⃣ DECAF — Decomposed Elementary Charts with Answerable Facts

• Focus: Factual lookup and comparative reasoning on simplified single-variable charts.
• Sources: Derived from ChartQA, ChartLlama, ChartInfo, DVQA.
• Content: 1,188 decomposed charts and 2,809 QA pairs.
• Tasks: Identify, compare, or extract values across clean, minimal visuals.

---

2️⃣ SPECTRA — Synthetic Plots for Event-based Correlated Trend Reasoning and Analysis

• Focus: Trend correlation and scenario-based inference between synthetic chart pairs.
• Construction: Generated via Gemini 1.5 Pro + human validation to preserve shared axes and realism.
• Content: 870 unique charts, 1,717 QA pairs across 333 contexts.
• Tasks: Analyze multi-variable relationships, infer trends, and reason about co-evolving variables.

---

3️⃣ STORM — Sequential Temporal Reasoning Over Real-world Multi-domain Charts

• Focus: Multi-step reasoning, temporal analysis, and semantic alignment across real-world charts.
• Source: Curated from Our World in Data with metadata-driven semantic pairing.
• Content: 648 charts across 324 validated contexts, 768 QA pairs.
• Tasks: Align mismatched domains, estimate ranges, and reason about evolving trends.

---

⚙️ Evaluation & Methodology

INTERCHART supports both visual and table-based evaluation modes.

• Visual Inputs:

  • Combined: Charts stitched into a unified image.
  • Interleaved: Charts provided sequentially.

• Structured Table Inputs:
  Models can extract tables using tools like DePlot or Gemini Title Extraction, followed by table-based QA.

• Prompting Strategies:

  • Zero-Shot
  • Zero-Shot Chain-of-Thought (CoT)
  • Few-Shot CoT with Directives (CoTD)

• Evaluation Pipeline:
  Multi-LLM semantic judging (Gemini 1.5 Flash, Phi-4, Qwen2.5) with majority voting to evaluate semantic correctness.

---

📊 Dataset Statistics

Subset	Charts	Contexts	QA Pairs	Reasoning Type Examples
DECAF	1,188	355	2,809	Factual lookup, comparison
SPECTRA	870	333	1,717	Trend correlation, event reasoning
STORM	648	324	768	Temporal reasoning, abstract numerical inference
Total	2,706	1,012	5,214	—

---

🚀 Usage

🔐 Access & Download Instructions

Use an access token as your Git credential when cloning or pushing to the repository.

1. Install Git LFS
   Download and install from https://git-lfs.com.
   Then run:

git lfs install

2. Clone the dataset repository
   When prompted for a password, use your Hugging Face access token with write permissions.
   You can generate one here: https://huggingface.co/settings/tokens

git clone [https://huggingface.co/datasets/interchart/Interchart](https://huggingface.co/datasets/interchart/Interchart)

3. Clone without large files (LFS pointers only)
   If you only want lightweight clones without downloading all image data:

GIT_LFS_SKIP_SMUDGE=1 git clone [https://huggingface.co/datasets/interchart/Interchart](https://huggingface.co/datasets/interchart/Interchart)

4. Alternative: use the Hugging Face CLI Make sure the CLI is installed:

pip install -U "huggingface_hub[cli]"

Then download directly:

hf download interchart/Interchart --repo-type=dataset

---

🔍 Citation

If you use this dataset, please cite:

@article{iyengar2025interchart,
  title={INTERCHART: Benchmarking Visual Reasoning Across Decomposed and Distributed Chart Information},
  author={Anirudh Iyengar Kaniyar Narayana Iyengar and Srija Mukhopadhyay and Adnan Qidwai and Shubhankar Singh and Dan Roth and Vivek Gupta},
  journal={arXiv preprint arXiv:2508.07630},
  year={2025}
}

---

🔗 Links

• 📘 Paper: arXiv:2508.07630v1
• 🌐 Website: https://coral-lab-asu.github.io/interchart/
• 🧠 Explore Dataset: Interactive Evaluation Portal