FysicsWorld: A Unified Full-Modality Benchmark for Any-to-Any Understanding, Generation, and Reasoning

[🏠 Project Page] [📖 Paper] [🤗 Dataset] [👾 ModelScope] [🏆 Leaderboard] [🀄 中文版]

🚀 News

• 2025.12.14 We release FysicsWorld, the first unified full-modality benchmark that supports bidirectional input–output across image, video, audio, and text, enabling comprehensive any-to-any evaluation across understanding, generation, and reasoning.

🎯 FysicsWorld Overview

We introduce FysicsWorld, the first unified full-modality benchmark that supports bidirectional input–output across image, video, audio, and text, enabling comprehensive any-to-any evaluation across understanding, generation, and reasoning. Our systematic design spans uni-modal perception tasks to fusion-dependent reasoning under strong cross-modal coupling, allowing us to diagnose, with unprecedented clarity, the limitations and emerging strengths of modern multimodal and omni-modal architectures. In contrast to existing omni-modal and multi-modal benchmarks, our FysicsWorld has several advantages:

• Diversity and High Quality. FysicsWorld is characterized by 8 "multi" properties, reflecting its comprehensive coverage, diversity, and robustness, namely: multi-dimensional (understanding, generation, reasoning, voice interaction), multi-modal (text, image, video, audio as both inputs and outputs), multi-task (16 primary tasks, 200+ sub-tasks), multi-source (3,268 samples from 40+ data sources and curated web data), multi-domain (170+ fine-grained open-domain categories), multi-type (closed-ended, open-ended, multiple-choice question, and image/video/audio generation), multi-target (evaluates Omni-LLMs, MLLMs, modality-specific models, unified understanding–generation models), and multi-assurance (multi-stage quality control strategies).

• Fusion-Dependent Cross-Modal Reasoning. We propose a method for omni-modal data construction, which is named Cross-Modal Complementarity Screening (CMCS) strategy, which ensures that our tasks maintain strong cross-modal coupling, preventing single-modality shortcuts and enforcing true synergistic perception of omni-modality.

• Speech-Driven Cross-Modal Interaction. To support natural, multimodal communication and interaction, we develop a speech-grounded multimodal data construction pipeline that ensures both linguistic fluency and semantic fidelity in voice-based interactions, including 10+ authentic voices and tones.

Based on FysicsWorld, we extensively evaluate various advanced models, including Omni-LLMs, MLLMs, modality-specific models, and unified understanding–generation models. By establishing a unified benchmark and highlighting key capability gaps, FysicsWorld provides not only a foundation for evaluating emerging multimodal systems but also a roadmap for the next generation of full-modality architectures capable of genuinely holistic perception, reasoning, and interaction.

🔍 Dataset Download

The full dataset, including associated multimedia files (images, videos, and audio), can be downloaded from:

• Link-1（🤗 HuggingFace）：[Link]
• Link-2（🤗 HF-Mirror）：[Link]
• Link-3（👾 ModelScope）：[Link]

🔮 Evaluation

To ensure a fair and standardized evaluation protocol, we release the full FysicsWorld dataset with ground-truth answers withheld, along with a test-mini subset (300 samples) that includes answers for local validation and debugging. You can find the QA data in ./data (full FysicsWorld) and ./test-mini (test-mini), respectively.

🕹️ Usage:

1. Download the full FysicsWorld dataset from here.
2. Run inference using your model on the provided questions.
3. Follow the guidelines, and format the model responses according to the required submission format.
4. Send the formatted responses to dicken@fyscis.ai. We will periodically update the corresponding scores on the leaderboard.

📈 Experimental Results

• Evaluation results of Omni-LLMs and proprietary MLLMs on image-centric omni-modal tasks

Task abbreviations: Task1-1 (Image Understanding), Task2-1 (Speech-Driven Image Understanding), Task2-2 (Image–Audio Contextual Reasoning), Task2-3 (Speech-Based QA on Image Content), Task2-4 (Speech Generation from a Person in an Image), and Task2-5 (Audio Matching from Image Context).

• Evaluation results of Omni-LLMs and proprietary MLLMs on video-centric omni-modal tasks.

Task abbreviations: Task1-2 (Video Understanding), Task3-1 (Speech-Driven Video Understanding), Task3-2 (Video–Audio Contextual Reasoning), Task3-3 (Speech-Based QA on Video Content), Task3-4 (Speech Generation from a Person in an Video), Task3-5 (Audio Matching from Video Context), and Task3-6 (Next-Action Prediction from Video Sequences and Current Visual State).

• Evaluation results of open-source MLLMs on modality-supported tasks.

Task abbreviations: Task1-1 (Image Understanding), Task1-2 (Video Understanding), and Task3-6 (Next-Action Prediction from Video Sequences and Current Visual State).

• Evaluation results of various models on (a) Audio Reasoning and (b) Video Generation.

📖 Citation

If you find FysicsWorld helpful for your research, please consider citing our work. Thanks!

@article{jiang2025fysicsworld,
    title={FysicsWorld: A Unified Full-Modality Benchmark for Any-to-Any Understanding, Generation, and Reasoning},
    author={Jiang, Yue and Yang, Dingkang and Han, Minghao and Han, Jinghang and Chen, Zizhi and Liu, Yizhou and Li, Mingcheng and Zhai, Peng and Zhang, Lihua},
    journal={arXiv preprint arXiv:2512.12756},
    year={2025}
}