๐Ÿง  EgoMemReason โ€” Leaderboard

Memory-driven reasoning over week-long egocentric video. 500 MCQs ยท Entity / Event / Behavior memory.

EgoMemReason

A Memory-driven Reasoning Benchmark for Long-Horizon Egocentric Video Understanding.

EgoMemReason is a 500-question multiple-choice benchmark over week-long egocentric videos (built on EgoLife). Models must answer questions whose evidence is sparsely distributed across hours or days, exercising three memory types:

  • Entity memory โ€” Cumulative State Tracking, Temporal Counting
  • Event memory โ€” Event Ordering, Event Linking
  • Behavior memory โ€” Spatial Preference Inference, Activity Pattern Inference

500 Qs ยท avg. 5.1 evidence segments / Q ยท avg. 25.9 h memory backtracking. The strongest model in the paper reaches 39.6% Overall.

Resources

Submission

Upload a JSON file with 500 entries:

[
  {"example_id": 1, "predicted_answer": "A"},
  ...
]

Questions have 4-10 options (letters A-J) โ€” predicted_answer must be a letter that appears in that question's options dict. See SUBMISSION_FORMAT.md for the full spec.

License

  • Annotations (this Space + the public dataset): CC BY-NC 4.0.
  • Video frames: governed by the EgoLife data license โ€” you must accept their terms separately.

Citation

@misc{wang2026egomemreasonmemorydrivenreasoningbenchmark,
      title={EgoMemReason: A Memory-Driven Reasoning Benchmark for Long-Horizon Egocentric Video Understanding},
      author={Ziyang Wang and Yue Zhang and Shoubin Yu and Ce Zhang and Zengqi Zhao and Jaehong Yoon and Hyunji Lee and Gedas Bertasius and Mohit Bansal},
      year={2026},
      eprint={2605.09874},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.09874},
}