ROAMM: A Benchmark Dataset for Multimodal Human Attention Decoding and EEG-to-Text Modeling During Naturalistic Reading
Haorui Sun, Ardyn Olszko, Niharika Singh, David Jangraw
Abstract
We present Reading Observed At Mindless Moments (ROAMM), a large-scale multimodal dataset comprising 50 hours of simultaneous EEG and eye-tracking recorded during naturalistic multi-page reading from 44 participants, with annotations including eye events, page-level comprehension scores, and word-level mind-wandering (MW) labels obtained via a retrospective self-report paradigm. We introduce a standardized evaluation protocol for MW detection under leave-one-subject-out evaluation, achieving up to 0.609 AUROC using supervised models. We also report results for EEG-to-text decoding trained on non-MW segments and show that decoding performance decreases when MW-labeled segments are included. Overall, ROAMM provides a benchmark dataset for MW detection and EEG-to-text decoding tasks, and enables the study of attention-related degradation in language decoding from brain activity in naturalistic reading.
BibTeX
@inproceedings{
sun2026roamm,
title={{ROAMM}: A Benchmark Dataset for Multimodal Human Attention Decoding and {EEG}-to-Text Modeling During Naturalistic Reading},
author={Haorui Sun and Ardyn Vivienne Olszko and Niharika Singh and David C. Jangraw},
booktitle={Forty-third International Conference on Machine Learning},
year={2026},
url={https://openreview.net/forum?id=zqLPdt09fE}
}