ICM-Bench: Person-Level Identity Reasoning in Multimodal Agents with Long-Term Memory

Paper | Code and evaluation

ICM-Bench evaluates person-centered evidence retrieval and cross-time relation reasoning in long-term multimodal agents. It contains 839 synthetic video clips spanning approximately 141 minutes and 1,217 open-ended questions about six recurring adults in a one-year life album.

Dataset contents

Item	Count
Videos	839
Date-stamped memory clips	838
Calibration clips	1
Retained shots	1,958
Recurring adults	6
Open-ended questions	1,217
Identity Recall	400
Cross-Episode Identity Retrieval	500
Long-Term Identity Profile Inference	317

The complete questions and annotations are provided in annotations/qa_test.jsonl and annotations/qa_test.json. They are not reproduced in this Dataset Card.

Download

hf download ryanren0330/ICM-Bench \
  --repo-type dataset \
  --local-dir ICM-Bench

cd ICM-Bench
tar -xf videos.tar

The archive extracts 839 files under videos/, named clip_000.mp4 through clip_838.mp4.

Repository structure

ICM-Bench/
├── README.md
├── LICENSE
├── CITATION.cff
├── videos.tar
├── videos/
│   └── metadata.jsonl
├── annotations/
│   ├── qa_test.jsonl
│   ├── qa_test.json
│   ├── characters.json
│   ├── dataset_statistics.json
│   └── schema.json
├── resources/
│   ├── asr_transcripts/
│   └── transcripts_with_speakers/
├── scripts/
└── checksums/sha256.txt

qa_test.jsonl is the recommended annotation file. qa_test.json contains the same records as a JSON array. Timestamped transcripts are available for the 829 memory clips containing scripted dialogue. Files in resources/asr_transcripts/ omit speaker names, whereas files in resources/transcripts_with_speakers/ retain speaker labels for reference and analysis.

Load the annotations

import json
from pathlib import Path

root = Path("ICM-Bench")
with (root / "annotations" / "qa_test.jsonl").open() as f:
    questions = [json.loads(line) for line in f]

print(len(questions))  # 1217
print(questions[0].keys())

Evaluation protocol

ICM-Bench is an evaluation-only benchmark with one test split. For Identity Recall and Cross-Episode Identity Retrieval, systems may access memory clips up to and including the question-specific before_clip. Long-Term Identity Profile Inference uses the complete timeline. Video-based settings additionally receive the calibration clip; transcript-only controls use speakerless transcripts.

The following evaluator-side fields must not be exposed to the evaluated system: reference_answer, target_character_ids, and evidence_video_ids. All evidence annotations use public clip-level identifiers from clip_000 to clip_838. Answers are open-ended and are evaluated by semantic equivalence rather than exact string matching.

The first video, clip_000.mp4, is a 24-second non-evidence calibration clip in which each recurring adult speaks a neutral sentence without revealing a name. It provides a shared face--voice reference before the 838 date-stamped memory clips.

Uses

Intended use

ICM-Bench is intended for evaluating long-term multimodal memory, person-centered retrieval, cross-episode reasoning, identity-profile inference, and open-ended video question answering.

Out-of-scope use

ICM-Bench is synthetic and should not be treated as a substitute for real-world video containing natural noise, occlusion, overlapping speech, diverse environments, and spontaneous social behavior. All depicted identities and voices are synthetic. The dataset must not be used to identify, track, profile, or make decisions about real people.

License

ICM-Bench is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. The videos were generated with Google Gemini/Veo services. Users remain responsible for complying with applicable law, provider terms, and the dataset license.

Code and paper

• Code and evaluation instructions: https://github.com/Shidu-Ren/ICM-Bench
• Paper: https://arxiv.org/abs/2609.04438

Citation

@article{ren2026icmbench,
  title={ICM-Bench: Person-Level Identity Reasoning in Multimodal Agents with Long-Term Memory},
  author={Ren, Shidu and Liu, Yunze and Liu, Xing and Wu, Chi-Hao and Zhou, Enmin and Shen, Junxiao},
  year={2026},
  journal={arXiv preprint arXiv:2609.04438},
  url={https://arxiv.org/abs/2609.04438}
}