--- license: apache-2.0 language: - en - zh tags: - omni - vlm - llm - reasoning - benchmark size_categories: - 1K

FysicsReason: Benchmarking Verifiable World-State Reasoning Across Omni-Modalities

🏠 Project Page    📖 Paper    🤗 Dataset    ## Dataset Overview FysicsReason is a five-task omni-modal benchmark designed to evaluate physical reasoning across images, audio, video, and text. The benchmark contains five complementary tasks that cover visual understanding, audio-visual grounding, temporal reasoning, physical property comparison, and quantitative physical inference. Each task is provided as a separate Parquet file under `task{1..5}/`, together with its associated media files. ## Task Composition | Task | Samples | Input | Prediction Target | | ----- | ------: | -------------------------------------------------- | ---------------------------------------- | | task1 | 562 | Image and text question | Target quantity and final answer | | task2 | 779 | Image, scene audio, and text question | Object bounding box and answer option | | task3 | 250 | Video and multiple-choice text question | Evidence interval and answer option | | task4 | 350 | Two object images, object audio, and text question | Property, materials, and selected object | | task5 | 387 | Video and text question | Physical property and numerical answer | ## Data Organization Media paths stored in the Parquet files are relative to the dataset root. For example: `data/task2/media/...` When loading the dataset locally, these paths should therefore be resolved relative to the directory containing the `data/` folder. Each task directory contains its corresponding annotation file and linked media resources, allowing the benchmark to be used directly for task-specific or unified omni-modal evaluation. ## Sample Identification The `index` field serves as the stable, zero-based row identifier within each task. For evaluation and result submission, a sample can be uniquely identified using the combination of: - `task_source` - `index` This convention provides a consistent mapping between dataset samples and inference outputs. ## Evaluation Inference results produced on FysicsReason can be evaluated using the official workflow provided in the FysicsReason repository. [FysicsReason Evaluation Repository](https://github.com/Fysics-AI/FysicsReason)