| language: | |
| - en | |
| license: cc-by-4.0 | |
| size_categories: | |
| - n<1K | |
| pretty_name: MIST | |
| tags: | |
| - video | |
| - benchmark | |
| - world-model | |
| task_categories: | |
| - other | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: test | |
| path: data/mist_cards.jsonl | |
| # MIST | |
| 43 Internal MIST case cards for delayed reveal. Each row names the scene setup, the hidden operation, the reveal query, and the state the last frame must show. | |
| A low memory score is often a missed action. The scene can already be set, and the hidden interval can stay leak-free, while the asked change is never carried out. These cards are the task record for that split: setup, the hidden change, and reveal. | |
| M3-Agent is a test-time rewrite of the missed move. It does not add a trained memory module. Keep-best is an oracle over the same score that reports the result. | |
| ## Load | |
| ```python | |
| from datasets import load_dataset | |
| cards = load_dataset("Wayne-King/MIST", split="test") | |
| ``` | |
| `data/mist_cards.json` is the same table in the `cases` shape read by the public control loop. | |
| ## Links | |
| - Paper: https://arxiv.org/abs/2610.23542 | |
| - Hugging Face paper page: https://huggingface.co/papers/2610.23542 | |
| - Project page: https://hku-mmlab.github.io/MIST/ | |
| - Code: https://github.com/HKU-MMLab/MIST | |
| ## Fields | |
| | Field | Meaning | | |
| | --- | --- | | |
| | `id` | Case id | | |
| | `category` | `ISM-R0` content, `ISM-R1` binding, or `ISM-R2` operation history | | |
| | `setup` | Scene before the change | | |
| | `hidden_operation` | The change that must happen while it stays hidden | | |
| | `reveal_query` | The later request to show the result | | |
| | `expected_reveal` | What the last frame must show | | |
| | `expected_present` / `expected_absent` / `expected_state` / `expected_count` | Scoring checks | | |
| The reveal text does not name the hidden answer as a gold sentence. `expected_*` is the scoring key. |