Video-MME-Logical

A Controlled Diagnostic Benchmark for Video Temporal-Logical Reasoning

Hohin Kwan*, Hongyu Li*, Ray Zhang, Manyuan Zhang, Xianghao Kong, Anyi Rao, Jiahao Xie, and Si Liu

* Equal contribution.

Project Page | Paper | Code | HF Paper

Abstract

Video-MME-Logical is a controlled benchmark for video temporal-logical reasoning: the ability to maintain, update, and compose evidence as visual states evolve across frames. It organizes evaluation around five temporal-logical operations and uses programmatic generation to control object states, transitions, temporal dependencies, and logical compositions. The benchmark supports difficulty-controlled final-answer evaluation and intermediate-state diagnostics that verify whether a model recovers the required reasoning trace before producing its final answer.

Experiments with state-of-the-art multimodal large language models reveal a substantial human-model gap, especially as temporal-logical complexity increases. Video-MME-Logical provides a scalable testbed for analyzing and improving this capability.

Benchmark at a Glance

Scope	Scale	Notes
Full benchmark described in the paper	503,750 videos	500,000 training videos and 3,750 test videos
Current Hugging Face release	3,750 test samples	3,750 videos and 900 referenced images
Task taxonomy	25 task categories	Five temporal-logical operations
Difficulty settings	3 levels	Easy, medium, and hard
Intermediate-state subset	8 task categories	Structured reasoning traces with exact-match verification

Release scope: this repository currently provides the 3,750-example test set. The 500K training videos described in the paper are not included in the current download.

Task Taxonomy

Operation	What it evaluates
State Tracking	Maintaining hidden or latent object states across visual transformations
Sequential Counting	Accumulating discrete evidence over time
Temporal Ordering	Recovering the order of state changes, revealed symbols, or event sequences
Dynamic Spatiality	Geometric and motion-based inference
Structural Composition	Composing spatial structures across viewpoints, occlusions, and partial observations

Easy, medium, and hard settings increase the temporal horizon and reasoning complexity while preserving the underlying task definition.

Controlled Construction

Each task category is implemented as an executable program with temporal transitions, scene configuration, metadata construction, and video rendering. Program-recorded metadata supports question construction, exact answer computation, difficulty control, and intermediate-state supervision.

Released Files

video_mme_logical.zip
`-- video-mme-logical/
    |-- three_level_testset.json
    |-- videos/                  # 3,750 video files
    `-- images/                  # 900 referenced images

Each record in three_level_testset.json contains:

Field	Description
id	Unique sample identifier
parent_major	High-level task family
major	Fine-grained task category
difficulty	Easy, medium, or hard setting
video_path	Relative path to the sample video
image_paths	Optional relative paths to referenced images
question	Evaluation prompt
answer	Ground-truth answer

Relative media paths are resolved from the directory containing three_level_testset.json.

Download

Install the current Hugging Face CLI and download the archive:

python3 -m pip install -U huggingface_hub
mkdir -p data/hf
hf download marcuskwan/video-mme-logical video_mme_logical.zip \
  --type dataset \
  --local-dir data/hf
unzip data/hf/video_mme_logical.zip -d data

After extraction, the test manifest is available at:

data/video-mme-logical/three_level_testset.json

Evaluation

The public GitHub repository includes a minimal Gemini API evaluator and a complete reproduction guide:

• Evaluation script
• Reproduction agent guide

Run a five-example smoke test before a full evaluation:

git clone https://github.com/Mrakas/video-mme-logical.git
cd video-mme-logical

python3 -m pip install -U huggingface_hub
mkdir -p data/hf
hf download marcuskwan/video-mme-logical video_mme_logical.zip \
  --type dataset \
  --local-dir data/hf
unzip data/hf/video_mme_logical.zip -d data

uv venv .venv
source .venv/bin/activate
uv pip install google-genai huggingface_hub

export GEMINI_API_KEY=your_key_here
python eval.py \
  --dataset data/video-mme-logical/three_level_testset.json \
  --model gemini-3-pro-preview \
  --limit 5 \
  --output runs/gemini-3-pro-preview_smoke_predictions.jsonl

The evaluator writes JSONL predictions and prints the final exact-match accuracy. API credentials are not included in this release.

Selected Zero-Shot Results

Accuracy (%) on Video-MME-Logical. The table is intentionally compact; see the paper or project leaderboard for full results.

Model	Overall	Easy	Medium	Hard
Human Level	95.9	98.4	95.9	93.4
Qwen2.5-VL-72B-Instruct	12.5	15.2	13.1	9.1
Qwen3-VL-30B-A3B-Think	10.3	16.0	8.7	6.1
GPT-5.4	22.7	31.7	20.3	16.1
Gemini-3.1 Pro	28.6	33.1	24.1	20.6

Intermediate-state evaluation is substantially harder: human performance is 96.1%, while GPT-5.4 and Gemini-3.1 Pro reach 17.4% and 10.8%, respectively, on Video-MME-Logical-S.

Citation

@article{kwan2026video,
  title={Video-MME-Logical: A Controlled Diagnostic Benchmark for Video Temporal-Logical Reasoning},
  author={Kwan, Hohin and Li, Hongyu and Zhang, Ray and Zhang, Manyuan and Kong, Xianghao and Rao, Anyi and Xie, Jiahao and Liu, Si},
  journal={arXiv preprint arXiv:2606.27828},
  year={2026}
}

License

The dataset release in this Hugging Face repository is distributed under the Creative Commons Attribution 4.0 International License. The evaluation code in the GitHub repository is distributed under the MIT License.