| license: apache-2.0 | |
| task_categories: | |
| - text-to-video | |
| # StateBench | |
| StateBench is a benchmark for world-state reasoning in video continuation, introduced in the paper [Do Video Generators Track the World Across Segments? A Benchmark and Method for World-State Reasoning in Video Continuation](https://huggingface.co/papers/2609.03673). | |
| Code: [https://github.com/AMAP-ML/StateAgent](https://github.com/AMAP-ML/StateAgent) | |
| The benchmark contains **200 cross-segment continuation tasks** across three difficulty levels: | |
| | Level | Count | Description | | |
| |-------|-------|-------------| | |
| | `past_visible` | 85 | The target object was visible in a prior frame — tests temporal memory. | | |
| | `occluded_process` | 65 | The state change occurs while the object is occluded — tests process reasoning. | | |
| | `complex_transition` | 50 | Multiple state transitions or conflicting cues — tests compositional reasoning. | | |
| Each task includes metadata such as `id`, `difficulty`, `target_object`, `shots`, `expected_state`, `checklist`, and `reference_frames`. |