Datasets:

mental-world-model
/

menti-bench

license: cc-by-nc-4.0
language:
  - en
task_categories:
  - question-answering
tags:
  - theory-of-mind
  - world-model
  - social-reasoning
  - mental-state
  - multimodal
size_categories:
  - n<1K
configs:
  - config_name: text
    data_files:
      - split: test
        path: text/text_instances.jsonl
  - config_name: image
    data_files:
      - split: test
        path: image/image_instances.jsonl
  - config_name: video
    data_files:
      - split: test
        path: video/video_instances.jsonl

Menti-Bench

Menti-Bench is a manually constructed, quality-controlled benchmark of situated decision scenarios for evaluating Mental World Modeling (MWM): whether a model can predict what a target agent will actually do next, in scenes where the correct prediction depends on tracking each agent's beliefs, knowledge access, goals, emotions, and social constraints rather than the physical scene alone.

Each instance presents a short story (text, an image sequence, or a sounding video), a designated target agent, a question about the target's next action, and six candidate actions. The correct action follows from the target's partial view of the scene, so distractor options are typically plausible under the global state but inconsistent with what the target can perceive, know, or socially do.

Paper: Mental World Modeling ([arXiv:XXXX.XXXXX])
Code (Mentis baseline system): https://github.com/mental-world/mentis

Composition

Config	Instances	Story carrier	Media
`text`	320	`story.text` narrative	—
`image`	100	ordered image sequence + `story.scene_context`	282 images (`image/assets/`)
`video`	28	sounding video + `story.scene_context`	28 mp4 clips (`video/assets/`)
total	448

Scenario coverage: interpersonal decisions (213), object/resource decisions (126), spatial/perceptual decisions (58), and risk/norm decisions (51), across everyday domains (home, workplace, school, public spaces, and more). Every instance is a single test split; the benchmark is evaluation-only and has no training split.

Fields

{
  "sample_id": "51",
  "modality": "image",
  "domain": "...", "domain_category": "...",
  "scene": "...", "scene_category": "interpersonal_decision",
  "num_of_characters": 3,
  "target_agent": "who the question is about, with identifying description",
  "question": "What will the target most plausibly do next?",
  "options": [
    {"option_id": "A", "action_description": "..."},
    {"option_id": "B", "action_description": "..."}
  ],
  "story": {
    "text": null,
    "images": ["image/assets/51_01.jpg", "image/assets/51_02.jpg"],
    "video": null,
    "scene_context": "identity anchors for the people shown in the media"
  },
  "answer": "C"
}

Exactly one of story.text, story.images, story.video is non-null. Media paths are relative to the dataset repository root. answer is the gold option id; intermediate annotations used during dataset construction (gold world states, gold observations, gold successor states) are not distributed, so the benchmark cannot be shortcut with oracle intermediate information.

Usage

from datasets import load_dataset
from huggingface_hub import snapshot_download

text = load_dataset("mental-world-model/menti-bench", "text", split="test")

root = snapshot_download("mental-world-model/menti-bench", repo_type="dataset")
image = load_dataset("mental-world-model/menti-bench", "image", split="test")
first_image_path = f"{root}/{image[0]['story']['images'][0]}"

To reproduce the paper's systems (direct baselines and the Mentis mental-world-model pipeline), run the accompanying code repository directly on the modality files:

python run.py predict --input <dataset-root>/text/text_instances.jsonl --system mentis

Construction and quality control

Stories, questions, and candidate actions were manually written and iteratively hardened so that wrong options each violate at least one constraint inferable from the scene (a character's belief or perceptual access, an object's state, an agreement, a norm, or timing). Images were synthetically generated, and videos were synthetically produced with spoken dialogue; both were manually quality-checked against their scripts. Gold answers were verified by multiple annotators, and a human reference score on the benchmark is reported in the paper.

Ethics and license

All scenarios, characters, and events are fictional; no real persons appear in any text, image, or video, and no personally identifiable information is included. The dataset probes everyday social reasoning and contains no graphic, hateful, or otherwise sensitive content; a small subset involves mild interpersonal conflict or safety-relevant workplace situations, which is the phenomenon under study.

Released under CC BY-NC 4.0 for non-commercial research use. Please do not use these instances for model training; the benchmark is intended as a held-out evaluation set, and training on it invalidates comparisons.

Citation

@article{mwm2026,
  title   = {Mental World Modeling},
  author  = {[AUTHORS]},
  journal = {arXiv preprint arXiv:[XXXX.XXXXX]},
  year    = {2026}
}