InSight: A Benchmark for Agentic Claim Verification in Interactive Visualizations

InSight evaluates multimodal agents on claim verification over interactive visualizations. Given an interactive HTML visualization and a natural-language proposition, an agent must explore the visualization over multiple turns (clicking, hovering, navigating) and classify the proposition as True, False, or NotEnoughInfo.

• Paper: arXiv:2609.01383
• Code / benchmark harness: github.com/maevehutch/insight

This repository contains the data only. The evaluation environment (Playwright-based browser interaction, model backends, prompts) lives in the GitHub repo.

Dataset structure

The dataset has two configs. Propositions reference visualizations by html_file; the visualizations themselves (~800 KB of HTML each) are stored once in their own config rather than duplicated into every row.

default config — propositions

Split	Rows	Description
test	500	The InSight test set, balanced across the three classes
extended	20,849	An extended pool of additional propositions over the same 297 visualizations

Field	Type	Description
source_id	string	Unique sample identifier
html_file	string	Key into the visualizations config, e.g. html/602.html
proposition	string	The statement to verify
label	int	Integer form of the gold label
class	string	Gold label: True, False, or NotEnoughInfo

visualizations config — interactive HTML

Split	Rows	Description
train	297	One row per visualization

Field	Type	Description
html_file	string	Path key matching html_file in the propositions
html	string	Full self-contained source of the interactive visualization

Raw files

The original files are also included so the benchmark runner can be used directly on a clone of this repo:

html/                       # 297 interactive HTML visualizations
raw/insight_test.jsonl
raw/insight_extended.jsonl

Usage

Load with datasets

from datasets import load_dataset

test = load_dataset("maevehutch/InSight", split="test")
viz  = load_dataset("maevehutch/InSight", "visualizations", split="train")
html_by_file = dict(zip(viz["html_file"], viz["html"]))

row = test[0]
print(row["proposition"], "->", row["class"])
html_source = html_by_file[row["html_file"]]   # render in a browser (e.g. Playwright) for the agent to interact with

Run the benchmark

The interaction loop, prompts, and model backends are in the GitHub repo. Clone both, then point the runner at the HF data:

git clone https://github.com/maevehutch/insight
git clone https://huggingface.co/datasets/maevehutch/InSight insight-data
cd insight && pip install -r requirements.txt && python -m playwright install chromium

python run_benchmark.py \
  --dataset ../insight-data/raw/insight_test.jsonl \
  --html-root ../insight-data \
  --model openai --model-name gpt-4o --api-key $OPENAI_API_KEY \
  --max-turns 10

See the GitHub README for all options, including local Hugging Face models and UI-TARS-style parsing.

License

The dataset is released under CC BY-NC-SA 4.0.

Citation

@article{insight2026,
  title   = {InSight: A Benchmark for Agentic Claim Verification in Interactive Visualizations},
  author  = {ADD AUTHORS},
  journal = {arXiv preprint arXiv:2609.01383},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.01383}
}