Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Server error while post-processing the rows. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

M³Eval: Multi-Modal Memory Evaluation through Cognitively-Grounded Video Tasks

Paper Project Page Hugging Face Dataset GitHub Code

Jie Huang1,*  Ruixun Liu1,*  Sirui Sun1  Xinyi Yang1  Yin Li2  Yixin Zhu1  Yiwu Zhong1,†

1Peking University  2University of Wisconsin-Madison

* Equal contribution. † Corresponding author.

News

  • 2026-11-16: update codebase.
  • 2026-6-4: We released the M³Eval benchmark, code, and project page.

M³Eval Overview

M³Eval overview

Abstract

As multi-modal models advance towards long-form video understanding, memory emerges as a critical capability. Despite substantial effort in developing video datasets and benchmarks, existing work primarily focuses on perception and reasoning, without systematically evaluating memory: what models retain, how faithfully information is preserved, and how robust memory remains under interference.

To address this gap, we introduce M³Eval, the first comprehensive evaluation framework and benchmark for probing different memory dimensions in multi-modal models. Grounded in cognitive psychology, our design features carefully constructed tasks isolating key aspects of memory. Leveraging M³Eval, we conduct extensive experiments across representative multi-modal models, revealing consistent weaknesses and distinctive behaviors.

We find that models struggle to maintain disentangled representations when processing parallel video streams, exhibit interference patterns differing substantially from those observed in human memory, ground memory sources more reliably in the spatial domain than the temporal domain, and demonstrate limited symbolic memory. Collectively, our benchmark provides a valuable resource for future research, whereas our findings highlight memory as a fundamental yet underexplored capability and offer insights for designing more effective memory mechanisms in multi-modal models.

Main Results

Divided Attention

Divided Attention main result

Accuracy (%) on three divided attention metrics under the split-screen setting without swaps and with frequent left/right swaps.

Memory Interference

Memory Interference main result

Proactive: first video (V1) interferes with recall of the second video (V2); retroactive: second video (V2) interferes with recall of the first video (V1). Delta denotes proactive minus retroactive.

Interleaved Events

Interleaved Events main result

Accuracy (%) on four interleaved reconstruction metrics.

N-Back

N-Back main result

Average accuracy under two symbolic attributes, scene and action, averaged over all K and N configurations.

Usage

Download the Dataset

The benchmark is composed from existing public benchmarks. Users must follow the original licenses and usage terms of each source dataset.

huggingface-cli download PKU-VaLuE-Lab/m3eval \
  --repo-type dataset \
  --local-dir data/m3eval

bash data/m3eval/unpack_archives.sh

After unpacking, data/m3eval/ contains the questions, cases, and videos used by the evaluation code. Each archive is independent and can be downloaded or resumed separately.

Dataset Examples

Divided Attention

Divided Attention dataset example

Simultaneous memory for two side-by-side videos.

Click to expand more examples

Memory Interference

Memory Interference dataset example

Interference between sequentially presented videos.

Interleaved Events

Interleaved Events dataset example

Memory reconstruction from temporally interleaved clips.

N-Back

N-Back dataset example

Decide whether the final clip matches the clip N positions earlier.

Citation

If you use M³Eval in your work, please cite:

@article{huang2026m3eval,
  title   = {M3Eval: Multi-Modal Memory Evaluation through Cognitively-Grounded Video Tasks},
  author  = {Huang, Jie and Liu, Ruixun and Sun, Sirui and Yang, Xinyi and Li, Yin and Zhu, Yixin and Zhong, Yiwu},
  journal = {arXiv preprint arXiv:2606.05008},
  year    = {2026}
}
Downloads last month
439

Paper for PKU-VaLuE-Lab/m3eval