WOVEN: Weaving Visual World Modeling into Multimodal LLMs

October 9, 2026: The training split is available now. We are uploading the validation split and the three test sets, and they will be available here soon.

WOVEN is a training source and benchmark for visual transition reasoning. A transition (s, a, s′) consists of the visual state of a scene before an action, the action, and the state after it. Visual transition reasoning infers the unobserved part of a transition from its observed parts: predicting an outcome, identifying the action behind a change, reasoning about an alternative action, or ordering intermediate states.

WOVEN organizes transition supervision by scene, action, and reasoning type, using diverse, realistic rollouts from a video generation model: 36,076 four-option multiple-choice items across 20 scene types, 5 action types, and 8 reasoning types. In the paper, training subsets of only about 2,000 WOVEN items each collectively improve 22 of 26 external benchmarks, by up to 27.3 percentage points.

Splits

Split	Items	Status
Training	22,728	Available
Validation	2,508	Uploading soon
In-distribution test	6,228	Uploading soon
Held-out-scene test (restaurant and beach scenes)	3,496	Uploading soon
State-perturbation test	1,116	Uploading soon
Total	36,076

Quick start

from datasets import load_dataset

ds = load_dataset("MLL-Lab/WOVEN", data_files="train/questions-*.parquet", split="train")

ex = ds[0]
print(ex["reasoning_type"], ex["action_type"], ex["scene"])
print(ex["question"])                          # images appear as <image_1>, <image_2>, ...
for opt in ex["options"]:
    print(f'{opt["label"]}. {opt["content"]}')  # image options are <image_k> as well
print("answer:", ex["answer"])
ex["images"][0]                                # PIL.Image

Turn an item into chat messages with the images in reading order:

import re

def to_messages(ex, instruction="Answer with the option's letter."):
    # The instruction is an example; the prompt templates used in the paper will be released with the code.
    text = ex["question"] + "\n" + "\n".join(f'{o["label"]}. {o["content"]}' for o in ex["options"])
    text += "\n" + instruction
    content = []
    for i, piece in enumerate(re.split(r"<image_(\d+)>", text)):
        if i % 2:
            content.append({"type": "image", "image": ex["images"][int(piece) - 1]})
        elif piece:
            content.append({"type": "text", "text": piece})
    return [{"role": "user", "content": content}]

Select items by their labels without decoding any images:

fwd = ds.filter(lambda t: t == "forward_dynamics", input_columns="reasoning_type")

Each training rollout is also available as 11 frames sampled every 0.5 seconds. Items and rollouts share source_id and action_id:

rollouts = load_dataset("MLL-Lab/WOVEN", data_files="rollout_frames/*.parquet", split="train")
r = rollouts[0]
print(r["source_id"], r["action_id"], r["action"])
r["frames"]                                    # 11 PIL images at r["timestamps"] = 0.0, 0.5, ..., 5.0 s

To download all files:

hf download MLL-Lab/WOVEN --repo-type dataset --local-dir WOVEN

Repository layout

train/questions-0000{0..4}-of-00005.parquet          22,728 training items, images embedded (2.5 GB)
rollout_frames/rollouts-0000{0..2}-of-00003.parquet  5,172 training rollouts, 11 frames each (1.5 GB)
assets/                                              figures used on this page

Data fields

Items (train/)

Field	Type	Description
id	string	Item identifier
source_id	string	Initial state; all rollouts from the same initial frame share it
action_id	string	The rollout the item is built from (action_0, action_1, ...)
action_type	string	exogenous (passive physical event), perceptive (camera motion), inspective (object inspection), navigative (navigation), or manipulative (object manipulation)
reasoning_type	string	One of the eight reasoning types below
scene	string	One of the 20 scene types
scene_group	string	indoors or outdoors
physical_principle	string	For passive physical events: permanence, cohesion, solidity, gravity, support, inertia, collision, or containment; otherwise null
manipulation_subject	string	For manipulation: human or humanoid; otherwise null
manipulation_view	string	For manipulation: egocentric or allocentric; otherwise null
question	string	Question text; images are referenced as <image_k>
options	list	Four options, each with label (A to D), content (text, or <image_k> for an image option), is_image, is_correct, distractor_type (the error type of a wrong option; correct for the answer), action_id (the rollout an option comes from; _static_ for the no-change option), order (the state order, for temporal ordering), and is_static (true for the no-change option)
answer	string	Letter of the correct option
images	list of images	All images the item shows, in order of first appearance (question, then options A to D); JPEG, 512 pixels wide

Rollouts (rollout_frames/)

Field	Type	Description
source_id, action_id	string	Keys shared with the items
action	string	Description of the action that produced the rollout
timestamps	list of floats	0.0, 0.5, ..., 5.0 seconds
frames	list of images	The 11 frames; JPEG, 512 pixels wide

Taxonomy

Reasoning types

Family	Reasoning type	reasoning_type	The question asks	Training items
Causal dynamics	Forward dynamics	forward_dynamics	What will the scene look like after the action?	4,132
Causal dynamics	Inverse dynamics	inverse_dynamics	Which action caused the change?	4,132
Counterfactual reasoning	Counterfactual removal	counterfactual_removal	What would the scene look like if the action had not happened?	2,060
Counterfactual reasoning	Counterfactual substitution	counterfactual_substitution	What would the outcome be under a different action?	2,060
Physical modeling	Outcome prediction	outcome_prediction	How will a passive physical event end?	1,040
Physical modeling	Cued prediction	cued_prediction	How will the event end, given a cue naming the physical principle?	1,040
Temporal coherence	Temporal ordering	temporal_ordering	In what order did these states occur?	4,132
Temporal coherence	Temporal adjacency	temporal_adjacency	Which states come right before and after a reference state?	4,132

Action types

action_type	Action type	Training items	Additional labels
exogenous	Passive physical events	4,160	physical_principle: permanence, cohesion, solidity (object properties); gravity, support, inertia, collision, containment (physical events)
perceptive	Camera motion	4,144
inspective	Object inspection	6,120
navigative	Navigation	4,144
manipulative	Object manipulation	4,160	manipulation_subject (human, humanoid) and manipulation_view (egocentric, allocentric)

Scenes

• Indoors: bathroom, bedroom, classroom, gym, hospital, kitchen, living room, office, restaurant, supermarket
• Outdoors: beach, construction site, crossroad, farm, garage, park, playground, school campus, sidewalk, yard

Restaurant and beach are held out: they appear only in the held-out-scene test.

How the data was built

• Rollouts are generated with a video generation model (Wan2.2-I2V-A14B) from specified initial frames and action descriptions. The resulting and intermediate states come from the rollout rather than from the conditioning prompt.
• Several rollouts under different actions start from the same initial state, so alternative outcomes of the same scene are available.
• Each rollout lasts 5 seconds, and 11 frames are sampled every 0.5 seconds; the questions are built from these frames.
• Each question is phrased with one of 30 templates drawn at random.
• The three wrong options are usually outcomes of other rollouts from the same initial state, and each is labeled with its error type. For physical-modeling items, the wrong options are produced by an image editor (Gemini 3.1 Flash Image) from a closed inventory of physically violating edit types.
• Images are stored as 512-pixel-wide JPEGs, the resolution used for training in the paper.

See the paper for the full construction and quality-control procedure.

Intended use

The training split is intended for training and analyzing visual transition reasoning in multimodal models. The test sets, once released, are intended for evaluation only; please do not train on them.

License

The dataset is released under CC BY 4.0.

Citation

@article{fan2026woven,
  title={WOVEN: Weaving Visual World Modeling into Multimodal LLMs},
  author={Fan, Zheyu and Zhang, Yue and Deng, Mingkai and Wang, Kangrui and Wang, Qineng and Chen, Canyu and Hao, Jie and Fan, Xing and Guo, Chenlei and Xing, Eric P. and Bansal, Mohit and Li, Manling},
  journal={arXiv preprint arXiv:2610.12417},
  year={2026}
}

For questions, please open a discussion on this page.