RiRi-y/unwritten-benchmark

The Unwritten Benchmark

Dataset Summary

The Unwritten Benchmark is a multimodal benchmark for acousto-kinematic word inference: given only the sound of a pen scratching on paper and/or the motion of a hand writing, models must infer the underlying word, even though no visible ink trace is present.

The dataset was introduced in the paper:

The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning
Garima Arya Yadav, Nilay Yilmaz, Yezhou Yang
Arizona State University

Paper: (https://riri-y.github.io/unwritten-benchmark/) (CVPR 2026 Findings)

The benchmark is designed to probe a capability that current multimodal systems still struggle with: recovering a symbolic outcome from the physical process that created it. Instead of recognizing explicit text or visible handwriting, models must reason from subtle motion and audio cues alone.

Supported Tasks and Leaderboards

This dataset supports research on:

• multimodal reasoning
• audio-visual inference
• handwriting understanding without visible text
• temporal perception and micro-kinematic reasoning
• cross-modal fusion under causal coupling

The main evaluation task is:

• Acousto-kinematic word inference: predict the written English word from one of three modalities:
  • Audio only
  • Muted video only
  • Audio + video

The primary metric used in the paper is:

• Ordered Letter Accuracy (OLA): the fraction of character positions correctly predicted in order

Languages

• English (en)

Dataset Structure

The benchmark contains samples spanning three handwriting styles:

• Standard: conventional non-connected print
• Cursive: flowing connected handwriting
• Retrace: a style with retraced strokes and repeated motion patterns

Each sample is provided in one or more of the following modalities:

• Audio: pen-on-paper sound only
• Muted Video: writing motion without sound
• Audio+Video: synchronized writing motion with sound

The paper describes two main data sources:

• Letter primitives: live-recorded individual letters used as the building blocks for synthesis
• Live Words: live-recorded full words for cross validation purposes
• Word-level samples:
  • semi-synthetic words created by concatenating style-consistent letter primitives
  • a smaller set of fully live-recorded words for validation and analysis

Data Collection Process

The data was collected from consenting adult participants. Three participants contributed handwriting in distinct styles:

• one participant for Standard writing
• one participant for Cursive writing
• one participant for a specialized Retrace style

Recording setup details from the paper:

• writing was performed with a dry pen on a blank white sheet of paper
• no visible ink trace was present in the recorded video
• audio and video were captured simultaneously
• audio-video pairs were manually synchronized to preserve temporal alignment

This design removes the explicit written trace and forces inference from indirect perceptual cues.

Example Use

Possible research uses include:

• benchmarking multimodal foundation models
• testing audio-visual fusion strategies
• studying causal reasoning from synchronized sensory signals
• evaluating handwriting inference without OCR-visible text
• comparing human and machine performance on dynamic perceptual tasks

Dataset Creation Rationale

This benchmark was created to measure a gap in current multimodal AI systems: the ability to infer a hidden symbolic outcome from a dynamic physical process.

Unlike OCR, handwriting recognition from visible traces, or standard audio-visual classification, this dataset removes the explicit target and requires reasoning from:

• hand motion
• stroke timing
• repeated or retraced movements
• pen-scratch rhythm and intensity

Privacy and Ethics

Ethics Statement: All data was collected from consenting adult participants who were informed of the study’s purpose. The dataset was fully anonymized, with no personally identifiable information collected or stored. The goal of this research is to advance scientific understanding of AI capabilities and limitations. We do not foresee any direct negative societal impacts, and the dataset does not contain any sensitive content.

Reproducibility Statement: The Unwritten Benchmark dataset has been made publicly available on our website: unwritten-benchmark.github.io. All models were accessed via their APIs or HuggingFace, and the exact model versions are noted in the main text. The prompts used for evaluation are provided in their entirety in the paper to ensure that our results are fully reproducible.

Citation

If you use this dataset, please cite:

@inproceedings{yadav2026unwritten,
  title={The Unwritten Benchmark: A New Challenge for Multimodal Machine Learning in Abstract Perceptual Reasoning},
  author={Yadav, Garima Arya and Yilmaz, Nilay and Yang, Yezhou},
  booktitle={CVPR Findings},
  year={2026}
}