# InfoVQA ![Category](https://img.shields.io/badge/Category-Reasoning-1f6feb) ![Metric](https://img.shields.io/badge/Metric-ANLS-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-Medium-d29922) ![Created](https://img.shields.io/badge/Created-Apr%202021-555) [![Paper](https://img.shields.io/badge/Paper-arXiv%202104.12756-b31b1b)](https://arxiv.org/abs/2104.12756) [![Dataset](https://img.shields.io/badge/Dataset-docvqa.org-0969da)](https://www.docvqa.org/datasets/infographicvqa) > **Quick answer:** InfoVQA (InfographicVQA, Mathew et al., 2021) tests question answering over dense, real-world infographics that combine text, charts, icons, and free-form graphical layouts — a harder document-understanding task than plain scanned-text documents. North Micro Vision scores 65.2% on the validation split as of August 2026. ## At a Glance **What it tests:** Question answering that requires reading and reasoning jointly over text, numerical data, and graphical elements (icons, maps, charts) laid out non-linearly across a single infographic image. **Why it matters:** Real-world documents like infographics, dashboards, and marketing materials mix free-form visual layout with text in ways plain-text OCR benchmarks don't capture; InfoVQA specifically probes this layout-aware reasoning. **Known limitations:** Many questions require multi-step numerical reasoning (e.g. summing values from a chart), so scores conflate visual layout understanding with arithmetic reasoning ability. ## What InfoVQA Measures InfoVQA collects real-world infographics — the kind found in reports, presentations, and public-facing communications — that densely combine running text, numerical tables, charts, icons, and non-linear visual layout. Unlike scanned-document benchmarks that test primarily text recognition, InfoVQA questions frequently require integrating information across multiple visual elements: comparing values in a chart, reading an icon's label, or combining a caption with a graphic. A meaningful share of questions also require basic numerical reasoning — such as computing a difference or a sum from values shown in the infographic — layering arithmetic ability on top of visual-layout and OCR competence. Because it demands both accurate text/number extraction and correct interpretation of non-standard visual layouts, InfoVQA sits at a harder point on the document-understanding spectrum than plain OCR or scanned-document QA benchmarks like DocVQA. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Reasoning (document/infographic understanding) | | Metric | ANLS (Average Normalized Levenshtein Similarity) | | Saturation | Medium | | Created by | Mathew et al. | | Source paper | [Mathew et al. 2021](https://arxiv.org/abs/2104.12756) | | Dataset | [docvqa.org InfographicVQA](https://www.docvqa.org/datasets/infographicvqa) | ## How InfoVQA Is Scored Answers are scored using ANLS (Average Normalized Levenshtein Similarity), which gives partial credit for near-matches to the ground-truth answer string rather than requiring an exact match — standard practice for open-ended document VQA where minor formatting differences shouldn't be penalized as full misses. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [North Micro Vision Instruct](/models/cohere-labs/north-micro-vision-instruct) | 65.2% | [North Micro Vision launch blog](https://huggingface.co/blog/CohereLabs/meet-north-micro-vision-instruct) | 2026-08 | *Score sourced from Cohere Labs' North Micro Vision Instruct launch announcement, August 2026, evaluated on the validation split.* ## InfoVQA on Benchgen > No Benchgen results yet — [be the first to run InfoVQA](/benchmarks/mathew-et-al/infovqa). ## InfoVQA vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **InfoVQA** | Dense infographics combining text, charts, and layout | Medium | | [DocVQA](/benchmarks/cvit-iiit-hyderabad/docvqa) | Scanned/plain-text document question answering | High | | [ChartQA](/benchmarks/vis-nlp/chartqa) | Chart-specific reasoning and value extraction | Medium | InfoVQA is a harder, more visually-complex sibling to DocVQA — use it when you need to evaluate a model's ability to parse mixed text-and-graphics layouts rather than mostly-linear scanned text. ## Run InfoVQA on Your Model Benchgen lets you evaluate your own vision-language model's infographic and document understanding and track results across model versions. ## Related Models & Benchmarks - Top performers: [North Micro Vision Instruct](/models/cohere-labs/north-micro-vision-instruct) - Related benchmarks: [DocVQA](/benchmarks/cvit-iiit-hyderabad/docvqa) · [ChartQA](/benchmarks/vis-nlp/chartqa) ## Frequently Asked Questions

What is InfoVQA?

InfoVQA (InfographicVQA) is a benchmark testing question answering over real-world infographics that combine text, charts, icons, and free-form visual layout.

What does a good InfoVQA score look like?

North Micro Vision Instruct reports 65.2% as of August 2026; strong document-focused vision-language models can reach 75-85%+.

Who created InfoVQA?

InfoVQA was created by Mathew et al.; see the [original paper](https://arxiv.org/abs/2104.12756).

Is InfoVQA saturated?

It shows medium saturation — frontier document-specialist models score highly, but it remains discriminative for general-purpose and compact vision-language models.

Rank	Model	Score
1	lfm3-vl-2b	86.2
2	lfm2-5-vl-3b	70.2
3	north-micro-vision-instruct	65.2

InfoVQA — Results

InfoVQA — Evaluations