FlowBench

FlowBench is a benchmark and training set for counterfactual spatial reasoning in vision-language models. Each question shows a scene, says how one object will move (for example "0.60 m to the right" or "about 8 cm up"), and asks about the scene after the move: which object will it be closest to, will it end up left of a landmark, will it hide another object, will it hit a wall. To answer, a model has to imagine how the scene changes, not just read the image.

Every question comes with the recorded after image, so the same question can also be asked as a plain perception question on the outcome. Answers are computed from geometry: simulator state for FlowBench-Sim and measured 3D geometry for FlowBench-Real. No model chooses the answers.

subset	scenes	split	questions	image pairs	image styles
FlowBench-Sim	AI2-THOR indoor houses	train (ProcTHOR-10k)	419,612	117,604 scene samples	natural, colour_som, red_som
		test (iTHOR floor plans)	714	226 scene samples
FlowBench-Real	real robot-manipulation episodes	train	8,575	5,367	colour_som, red_som
	(Bridge V2, DROID, Language-Table, RT-1, Berkeley UR5)	test	1,083	1,083

In both subsets, train and test come from disjoint scenes: different houses in Sim, different image pairs in Real.

Contents

• Task
• Configurations and loading
• Question categories
• Image styles
• FlowBench-Sim
• FlowBench-Real
• Evaluation notes
• Limitations
• Licences
• Citation

Task

Each row has a before image image_before, an after image image_after, and two versions of the question:

field	shown with	what the model does	example (Sim, colour_som)
motion_command + question	image_before	counterfactual: predict what happens after the stated move	move the kettle (marked orange) 0.60m right + "After the move, which three objects are closest to the kettle (marked orange)?"
static_question	image_after	static: read the outcome off the after image	"Which three objects are closest to the kettle (marked orange)?"

Both versions share the options choice_A … choice_F and the answer answer. The gap between static and counterfactual accuracy shows how much a model loses when it has to imagine the outcome instead of seeing it.

A FlowBench-Real example (colour_som, RT-1):

For scale, the object marked yellow is about 6 cm across at its widest point. If the robot arm moves the object marked blue so that, from the camera's perspective, its center ends up about 14 cm to the left, and about 19 cm down from where it was before (directions as seen from the camera), which of these will the object marked blue end up closest to, by straight-line distance? A. The object marked yellow. B. The object marked red. C. The object marked cyan.

Configurations and loading

config	subset	image style	train	test	parquet size
sim_colour_som (default)	Sim	coloured marks	419,612	714	~62 GB
sim_red_som	Sim	red labelled badges	419,612	714	~61 GB
sim_natural	Sim	clean renders	419,612	714	~46 GB
real_colour_som	Real	coloured marks	8,575	1,083	~1.2 GB
real_red_som	Real	red numbered badges	8,575	1,083	~1.0 GB

Within a subset, all styles share question ids (id). In Sim they also share the option order and answer letter, except for some natural rows (see choices_match_som). In Real they share the options and answer letter.

from datasets import load_dataset

# test sets are small
sim_test = load_dataset("StanfordNeuroAILab/FlowBench", "sim_colour_som", split="test")
real_test = load_dataset("StanfordNeuroAILab/FlowBench", "real_colour_som", split="test")

row = sim_test[0]
row["image_before"], row["image_after"]  # PIL images
row["motion_command"], row["question"], row["answer"]

# Sim train is large; stream it
sim_train = load_dataset("StanfordNeuroAILab/FlowBench", "sim_colour_som", split="train", streaming=True)

Options are stored in separate columns, and unused letters are null. A simple way to build a prompt for Sim:

letters = [l for l in "ABCDEF" if row[f"choice_{l}"] is not None]
options = "\n".join(f"{l}. {row[f'choice_{l}']}" for l in letters)
counterfactual_prompt = f"{row['motion_command']}\n{row['question']}\n{options}"  # with image_before
static_prompt = f"{row['static_question']}\n{options}"                           # with image_after

FlowBench-Real rows also include the exact text used in our evaluations: prompt for the counterfactual question and static_prompt for the static one.

Question categories

category	question (about the scene after the move)	Sim train	Sim test	Real train	Real test
closer_to	Which of several objects will the moved object be closest to?	59,474	94	209	189
closest_three	Which three of five or six objects will be closest to it? (3-letter answer)	66,679	112	81	135
count_near	How many of the listed objects will be within a given distance of it?	65,908	110	317	131
left_right_landmark (Sim) / left_right (Real)	Will it be left or right of a landmark, as seen from the camera?	67,554	116	448	193
near_object	Will it be within a given distance of another object?	67,636	117	6,831	193
occlusion	Will a landmark be fully visible, or partly or fully hidden by it?	13,576	40	20	96
feasibility	Starting from the after state, can it make a further stated move without colliding?	45,869	106	–	–
contact_wall	Will it be touching a wall?	32,916	19	–	–
depth_order	Will it be closer to the camera than a landmark? (Bridge V2 and DROID only)	–	–	669	146
total		419,612	714	8,575	1,083

mcq_type gives the format: mcq2 (yes/no), mcq3, mcq4, mcq5_select3 or mcq6_select3. For count_near, the options are counts.

Image styles

To refer to specific objects, most styles draw marks on the image (Set-of-Marks, SoM):

style	image	how the text refers to objects
natural (Sim only)	clean render	names and positions: "the 2nd kettle from the left"
colour_som	solid coloured dots with a white rim and no text. Colours are sampled for each question; in Real, a dot never has the colour of the object under it	Sim: "the kettle (marked blue)"; Real: "the object marked orange"
red_som	red badges with a label	Sim: mover 1, landmark or occludee 2, option objects A–F ("the kettle (1)", "(A) the plate"). Real: numbers on every marked object ("the object marked 3")

In the after image, marks are drawn where the objects ended up. In FlowBench-Real, the 103 test and 681 train named-object questions (Bridge V2 and DROID, named_objects = true) have no marks. They refer to objects by name ("the leftmost soup can") and appear in both Real styles with clean frames.

---

FlowBench-Sim

Source and generation

Scenes come from AI2-THOR. Train uses ProcTHOR-10k houses and test uses the hand-built iTHOR floor plans, so no test house appears in training. In each scene sample (sample_id), one object (mover_object_id) is moved by a stated distance and direction in the camera frame ("left", "right", "forward along the camera axis", ...). The simulator renders the scene before and after the move. Each category has its own view (view_id = view_<category>, camera plus move) and one question per view. The answer is computed from simulator metadata: bounding-box distances, screen positions, occlusion cover, collision-checked probes and wall clearances.

Filters on the released rows:

• ground truth is complete;
• the mover is not a structural object (windows, doors, cabinets, drawers, sinks, faucets, countertops);
• every referenced mark is drawn;
• for occlusion, the occludee covers at least 150 px, and any cover by the mover counts as occluded;
• static questions never mention the move;
• answer letters are balanced within each category. Yes/no questions are balanced within each polarity, so a letter never stands for "yes".

How the three styles relate

• colour_som is the reference version of every question.
• red_som asks the same questions with the same choice permutation. Mapping each red mark id to its colour gives the colour_som row exactly, which was checked for every row. In 66,601 train and 205 test rows the permutation moves lettered option objects, so the red badges were redrawn with the new letters (som_relabelled).
• natural comes from the unmarked export of the same scene samples, with the same text fixes. When it asks about the same objects as the SoM rows, its options follow the SoM order (309,142 of 419,612 train rows and 510 of 714 test rows; choices_match_som = true). The other rows, which are most closer_to rows and all count_near rows, use different distractors or object lists. They are separate questions with their own balanced letters.

Reasoning annotations (cot_gt, cot_1, cot_0)

Sim rows include up to three chains of thought, each ending with the answer. They can be used for supervised reasoning training.

field	written by	inputs
cot_gt	deterministic rules for each category	simulator metadata before and after the move (distances, screen positions, cover rates, collision checks, wall clearances)
cot_1	Qwen3-VL-32B-Instruct (most categories); Qwen3.5-9B (feasibility, contact_wall)	question, before and after images, cot_gt
cot_0	Qwen3.5-9B (thinking off)	question, motion command and the before image only, with cot_1 as a private hint

• cot_gt is written separately for each style, in that row's wording and options. It states the rule that produced the category's ground truth, the recorded numbers and the conclusion, and the conclusion matches the row's answer. It is present on 417,748 / 419,612 train rows and 711 / 714 test rows. It is null on 1,811 train and 3 test closest_three rows whose stored answer disagrees with their own after image; those answers are probably wrong. It is also null on 53 natural rows that list two options with the same text.
• cot_1 / cot_0 are train-only. They were written for the natural question and appear on 278,662 natural rows and on 248,777 rows each in colour_som and red_som, wherever the SoM question asks about the same objects. Each cot_0 passed automatic checks: its final answer equals cot_1's; it does not mention an after or second image or any hint; it never states a direction that contradicts the motion; its word overlap with cot_1 is at most 0.8; and it is 25–260 words long. Because they are VLM-written, these texts can get scene details wrong when those details do not affect the answer.

Sim row fields

field	description
id	<sample_id>__<question_id>, identical across the three styles
sample_id, scene, view_id, question_id	scene sample, AI2-THOR scene, view (camera and move), question
category, mcq_type, n_choices	question category, format, number of options
image_before, image_after	before and after images (800×600)
motion_command	the move, e.g. "move the bowl (marked blue) 0.50m left"
question, static_question	counterfactual and static question text
choice_A … choice_F, answer	options (null if unused) and the correct letter(s)
source_version	generator version of the scene sample (procthor_v9 / procthor_v10 / ithor_v9 / ithor_v10)
mover_object_id	AI2-THOR id of the moved object
internal_split	train only: the split used in our experiments (train0 137,684 · train1 137,599 · train2 137,612 · val 1,916 · test 4,801). It is a held-out part of the ProcTHOR train data, separate from the iTHOR test split
occludee_px, occlusion_cover, occlusion_relabelled	occlusion only: occludee pixels in the before image, fraction covered by the mover, whether the row was relabelled from "visible"
som_colors	colour_som only: JSON string mapping mark id → colour (1 mover, 2 landmark or occludee, A–F option objects)
som_labels, som_relabelled	red_som only: JSON string mapping mark id → label drawn on the badge; whether badges were redrawn with the release letters
choices_match_som	natural only: whether the options are the same objects in the same order as the SoM rows (if so, the answer letter matches too)
cot_gt, cot_1, cot_0	reasoning annotations (see above); null where not available

Image encoding. Sim test images are the original PNGs. Sim train images are JPEG at quality 95, which keeps the train parquet small.

---

FlowBench-Real

Source and generation

Questions come from real robot-manipulation episodes in Bridge V2, DROID, Language-Table, RT-1 and Berkeley UR5. The before image is the first frame of an episode. The after image is the last frame, or for DROID a frame after the last gripper release. The question states the moved object's measured displacement in centimetres, and the after frame is the recorded outcome.

split	questions	Language-Table	RT-1	DROID	Bridge V2	Berkeley UR5
train	8,575 (5,367 image pairs)	3,039	2,144	1,905	1,449	38
test	1,083 (1,083 image pairs, one question each)	478	91	245	267	2

Train and test share no image pair.

How the answers are computed.

1. SAM 3 segments both frames, Depth Anything 3 estimates depth for both frames in one call, and AllTracker tracks every pixel through the episode.
2. Objects are matched across frames by their tracks. The moved object must move at least 3 cm and landmarks less than 2 cm, and the tracked motion must agree with the measured displacement.
3. 3D positions are X = z / g · K⁻¹ [u, v, 1]ᵀ, where g is a divisor set for each dataset (Bridge V2 1.95, DROID 1.15, RT-1 0.61, Berkeley UR5 0.84). For Language-Table, positions come from a table-plane homography fitted to the robot's recorded end-effector positions.
4. Distances are between object surfaces (centre distance minus estimated radii), and every comparison must clear a margin of 3 cm (6 cm for DROID).
5. Rule-based filters remove ambiguous items. Most of these filters were added after a round of human review found a failure case.

A VLM only names objects and checks marks; it never selects an answer. In the test set, yes/no answers are 45–55% "yes" within each category. no_change marks the 291 test questions whose answer the move does not change; in all other test questions the answer differs between the before and after frames. 96 test questions were checked by a human reviewer against the after frame and confirmed correct (human_verified).

The motion text gives the displacement of the object's centre in the image's left/right and up/down directions, in whole centimetres. Components under 2 cm and the depth component are left out. Each question also includes a scale reference measured the same way ("For scale, X is about N cm across at its widest point").

Real row fields

field	description
id, sample_id	question id <source subset>__<number> (unique across splits and identical across the two styles); image-pair id
source, source_subset	robot dataset; generation pool
category, mcq_type, n_choices	question category, format, number of options
image_before, image_after	before and after frames (640×480; DROID 912×512), original JPEG bytes
scale_reference	scale sentence (null on 76 train rows)
motion_command	the move, e.g. "… its center ends up about 14 cm to the left, and about 19 cm down …"
question, static_question	counterfactual and static question text
choice_A … choice_F, answer	options (null if unused) and the correct letter(s)
prompt, static_prompt	full text used in our evaluations (scale sentence, move, question, lettered options)
named_objects	whether this is a named-object question (no marks)
no_change	whether the move leaves the answer unchanged
secondary_movers	JSON list of other objects the robot moved in the same episode, which the move text describes
provenance	JSON with the measured quantities behind the answer (distances before and after in metres, threshold, margin, depth gap, occlusion rule, ...)
som_colors	colour_som only: JSON string mapping mark number → dot colour (null in red_som and on named-object rows)
human_verified	test only: true for the 96 questions a human checked and confirmed; null otherwise

---

Evaluation notes

• Answers are option letters A–F. For closest_three the answer is three letters in alphabetical order (e.g. ABD).
• Report the counterfactual setting (before image with motion_command and question) and the static setting (after image with static_question) separately. Their difference measures how much the model loses when it must imagine the outcome.
• In FlowBench-Sim, contact_wall and near_object are usually left out of headline averages.
• Sim train includes internal_split. For held-out evaluation, use the test split (iTHOR scenes for Sim, disjoint image pairs for Real).

Limitations

• Templated language. The questions are templated by design to isolate spatial reasoning. They do not cover the variety of free-form instructions.
• Sim ground truth uses simulator geometry: bounding-box distances, AI2-THOR collision checks and wall clearances, and moves are stated in the camera frame.
• Real ground truth uses monocular depth, segmentation and tracking, so absolute centimetres are approximate. On Bridge V2 the size check is confounded by toy food. Relative geometry is reliable, and comparisons must clear a margin.
• Coverage. Real depth_order exists only for Bridge V2 and DROID. Language-Table and RT-1 have no named-object questions. Real train is dominated by near_object.
• Dense ground truth not included. The parquet files contain the images and question annotations. The dense ground truth used to build the questions (optical flow, depth, instance segmentation, moved-object masks, per-sample metadata) is not in the parquet files.

Licences

• FlowBench-Sim images are rendered with AI2-THOR using ProcTHOR-10k and iTHOR scenes.
• FlowBench-Real frames come from Bridge V2, DROID, Language-Table, RT-1 and Berkeley UR5 and keep their original licences. Check the source dataset's terms before redistributing frames.

Citation

@misc{flowbench,
  title  = {FlowBench},
  author = {TODO},
  year   = {2026},
  note   = {TODO: add paper reference}
}