sysuwyh357/ROSE-v0.1

ROSE: Benchmarking the Perception-to-Action Gap in Multimodal Models

ROSE (Reference-conditioned Oddity and Symbolic Execution) is a controlled benchmark for evaluating whether multimodal large language models can turn fine-grained visual evidence into the symbolic action required by the current task context.

• Paper: arXiv:2606.19965
• PDF: arXiv PDF
• Project page: https://xbdxwyh.github.io/ROSE-v0.1/
• Evaluation code: https://github.com/xbdxwyh/ROSE-v0.1
• Dataset: https://huggingface.co/datasets/sysuwyh357/ROSE-v0.1

Authors: Yihao Wang, Zijian He, Jie Ren, Keze Wang
Affiliations: Sun Yat-sen University; Shaanxi Normal University
Corresponding author: Keze Wang (kezewang@gmail.com)

Overview

MLLMs are increasingly expected not only to recognize visual content, but also to act on it. ROSE isolates this perception-to-action interface by holding the visual scene fixed while varying the relevant region constraint and the required symbolic output.

Each ROSE scene is a grid of visually similar elements with a small set of exceptions. The target identity is not named in the instruction. A model must infer the scene-internal majority reference, identify the exception cells, and then read out the same visual evidence under different contexts: global counting, local counting, local clicking, visual-region clicking, or exclusion-conditioned clicking with count submission.

This scene-coupled design makes it possible to ask whether a visual interpretation that supports a compact counting answer also supports exact, context-dependent coordinate action.

Dataset at a glance

Property	Value
Visual sources	5
Scenes	1,512
Rendered images	3,024
Task instances	7,560
Development instances	2,000
Test instances	5,560
Coupled task templates	5
Main output operations	COUNT, CLICK, SUBMIT

Each scene produces five coupled task instances and two renderings: an uncued base image and a cue-augmented image for visual-region clicking. The split is performed at the scene level, so all task variants and renderings derived from the same scene remain in the same split.

Visual sources and composition

Subset	Controlled visual source	Scenes	Images	Dev tasks	Test tasks
ROSE-ChineseGlyph	Confusable Chinese characters rendered with the same verified font	412	824	555	1,505
ROSE-EmojiStyle	Same emoji identity, different rendering providers	300	600	395	1,105
ROSE-EmojiContent	Related emoji identities rendered in a shared style	300	600	395	1,105
ROSE-PixelEdit	Same pixel-art asset with a localized edit	300	600	395	1,105
ROSE-PixelContent	Related but distinct pixel-art assets	200	400	260	740
Total		1,512	3,024	2,000	5,560

The five sources are designed to test whether the perception-to-action gap persists across qualitatively different fine-grained visual evidence rather than depending on a single category or cue.

Coupled task suite

Each scene is instantiated into five task templates. The exact task_template strings in the released metadata should be treated as authoritative.

Template	Paper abbreviation	Region/context	Required output
T1_COUNT_GLOBAL	G-Cnt	Whole grid	COUNT(n)
T2_COUNT_LOCAL_NUMERIC	L-Cnt	Numerically specified row range, column range, or rectangle	COUNT(n)
T3_CLICK_LOCAL_NUMERIC	L-Clk	Numerically specified row range, column range, or rectangle	CLICK(...); DONE
T4_CLICK_VISUAL_REGION	V-Clk	Highlighted region in the image	CLICK(...); DONE
T5_CLICK_EXCLUSION_SUBMIT	Excl-CS	Outside a text-specified excluded row, column, or rectangle	CLICK(...); SUBMIT(n)

T1 and T2 provide counting-oriented readouts. T3--T5 test whether the same scene-level evidence supports increasingly explicit and context-sensitive coordinate actions.

Formal output protocol

ROSE uses an exact symbolic output space.

Counting

COUNT(n)

Example:

COUNT(3)

Clicking

CLICK(Rr,Cc); CLICK(Rr,Cc); ...; DONE

Example:

CLICK(R2,C4); CLICK(R5,C1); DONE

When no target cell is valid under the current region:

DONE

Clicking with count submission

CLICK(Rr,Cc); CLICK(Rr,Cc); ...; SUBMIT(n)

Example:

CLICK(R1,C4); CLICK(R3,C2); SUBMIT(2)

When no target remains after exclusion:

SUBMIT(0)

Rows and columns are 1-indexed. Rows are counted from top to bottom, and columns from left to right. For click-based tasks, click order is ignored during evaluation, but the predicted coordinate set must match the ground-truth set exactly. Duplicate clicks, invalid coordinates, malformed outputs, and inconsistent click-count submissions are strict failures.

Repository structure

ROSE-v0.1/
├── README.md
├── metadata_dev.jsonl
├── metadata_test.jsonl
├── export_summary.json
└── images/
    ├── ROSE-ChineseGlyph/
    ├── ROSE-EmojiStyle/
    ├── ROSE-EmojiContent/
    ├── ROSE-PixelEdit/
    └── ROSE-PixelContent/

All physical images are stored once under the shared images/ directory. Development and test are logical task-instance splits represented by metadata_dev.jsonl and metadata_test.jsonl.

Multiple task rows may reference the same image_id and repository-relative image path. This is intentional: ROSE applies different task contexts and output operations to the same visual scene.

Loading the dataset

Install dependencies

pip install -U datasets huggingface_hub pillow

Load the metadata splits

from datasets import load_dataset

repo_id = "sysuwyh357/ROSE-v0.1"
ds = load_dataset(repo_id)

print(ds)
print(ds["dev"])
print(ds["test"])
print(ds["test"][0])

Expected split sizes:

DatasetDict({
    dev:  2000 rows
    test: 5560 rows
})

The image column stores a repository-relative path, such as:

images/ROSE-ChineseGlyph/example.png

Load one image on demand

from datasets import load_dataset
from huggingface_hub import hf_hub_download
from PIL import Image

repo_id = "sysuwyh357/ROSE-v0.1"
ds = load_dataset(repo_id)
row = ds["test"][0]

local_image_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=row["image"],
)

image = Image.open(local_image_path).convert("RGB")

print(row["item_id"])
print(row["prompt"])
print(row["answer"])
image.show()

Download the complete benchmark once

This is recommended for full-model evaluation because many task instances reuse the same images.

from pathlib import Path

from datasets import load_dataset
from huggingface_hub import snapshot_download
from PIL import Image

repo_id = "sysuwyh357/ROSE-v0.1"
local_root = Path(snapshot_download(repo_id=repo_id, repo_type="dataset"))
ds = load_dataset(repo_id)

row = ds["test"][0]
image = Image.open(local_root / row["image"]).convert("RGB")

print(row["prompt_global"])
print(row["prompt_local"])
print(row["ground_truth_formal_answer"])

Filter by subset or task

from datasets import load_dataset

repo_id = "sysuwyh357/ROSE-v0.1"
ds = load_dataset(repo_id)

test = ds["test"]

glyph = test.filter(
    lambda x: x["rose_subset"] == "ROSE-ChineseGlyph"
)

visual_region_click = test.filter(
    lambda x: x["task_template"] == "T4_CLICK_VISUAL_REGION"
)

print(len(glyph))
print(len(visual_region_click))

Stream the metadata

from datasets import load_dataset

repo_id = "sysuwyh357/ROSE-v0.1"
streamed = load_dataset(repo_id, split="test", streaming=True)

for row in streamed.take(3):
    print(row["item_id"], row["task_template"], row["image"])

Streaming loads metadata rows. Use hf_hub_download to retrieve image files when needed.

Metadata format

Each JSONL row corresponds to one task instance. The exported row preserves the original item metadata and adds canonical release fields for direct use.

Core identifiers

Field	Description
item_id	Unique task-instance identifier
scene_id	Identifier shared by all task variants derived from the same scene
image_id	Identifier of the rendered image used by the task
rose_subset	One of the five ROSE visual subsets
split	Official split: dev or test

Image fields

Field	Description
image	Repository-relative path to the released image
source_image_path	Original subset-relative path before release export, when available

Additional image-record fields may be exported with an image_ prefix when available.

Prompt fields

Field	Description
prompt_global	Shared protocol instruction
prompt_local	Task-specific instruction defining the current region or operation
prompt	Convenience field formed by joining prompt_global and prompt_local

The paper uses the following multimodal input order:

prompt_global → image → prompt_local

Users evaluating other models should preserve this order unless explicitly studying prompt-order sensitivity.

Task fields

Field	Description
task_template	Coupled task-template identifier
task_mode	COUNT, CLICK, or CLICK_COUNT
region_type	GLOBAL, ROW_RANGE, COL_RANGE, RECTANGLE, VISUAL_REGION, or EXCLUSION
region_params	Parameters defining the active or excluded region
region_case	Relation between the region and the global exception set, when available
cue_type	Visual cue type, such as no cue, outline box, or mask
difficulty	Dataset difficulty metadata, when available
expected_output_format	Expected formal output format

Ground-truth fields

Field	Description
ground_truth_count	Number of valid target cells under the current task context
ground_truth_clicks	Exact valid target coordinates for click-based tasks
ground_truth_formal_answer	Canonical formal answer
answer	Convenience alias for the canonical answer
all_target_cells	Global exception cells, when available
region_target_cells	Exception cells selected by the current region, when available

Scene fields

Common scene metadata is copied into each task row so that most analyses do not need to join a separate scene table.

Field	Description
rows, cols	Grid dimensions
cell_size	Rendered cell size
image_width, image_height	Rendered image dimensions, when available
grid_size_profile	Grid-size category, when available
global_target_count	Number of exceptions in the complete scene
target_cells	Global exception coordinates
primitive_type	Fine-grained visual source type

Subset-specific fields may also be present, including glyph identities and fonts, emoji identities and styles, visual-similarity metadata, and pixel-art pair information. Fields that are not applicable to a subset may be absent or null.

Schematic example

The following record is illustrative. Exact fields vary by task and subset.

{
  "item_id": "ITEM_000001",
  "scene_id": "SCENE_000001",
  "image_id": "IMAGE_000001",
  "rose_subset": "ROSE-ChineseGlyph",
  "split": "test",
  "image": "images/ROSE-ChineseGlyph/example.png",
  "task_template": "T3_CLICK_LOCAL_NUMERIC",
  "task_mode": "CLICK",
  "region_type": "RECTANGLE",
  "region_params": {
    "r1": 2,
    "r2": 5,
    "c1": 3,
    "c2": 7
  },
  "prompt_global": "...",
  "prompt_local": "...",
  "prompt": "...",
  "ground_truth_count": 1,
  "ground_truth_clicks": [
    "R4,C6"
  ],
  "ground_truth_formal_answer": "CLICK(R4,C6); DONE",
  "answer": "CLICK(R4,C6); DONE",
  "rows": 8,
  "cols": 10,
  "global_target_count": 3
}

Official splits

ROSE provides no training split.

• dev is intended for prompt validation, protocol debugging, and evaluator integration.
• test is the official benchmark split used for the reported evaluation.

The split is performed at the scene level. All task instances derived from the same scene remain in the same split, preventing closely related task variants from appearing across development and test.

Users should not merge the official splits when reporting benchmark results.

Evaluation

The primary metric is strict exact task success (PASS). A prediction passes only when it follows the required formal grammar and exactly satisfies the current task.

The evaluator also supports diagnostic metrics including:

Metric	Meaning
VALID	Grammar-valid output rate
SOFT	Partial credit based on count error or click-set quality
C-F1	Strict coordinate-set F1 over click-applicable tasks
R-OK	Region-rule compliance for predicted in-grid clicks
Card-Exact	Exact clicked-cardinality match
Action Ret.	Mean action PASS after conditioning on correct global counting

Evaluation code and model-format examples are available in the ROSE code repository.

Reported results

The paper evaluates nine recent MLLMs and one trained human annotator as a solvability reference. Human performance reaches 98.8% average PASS, while model performance ranges from 14.3% to 92.2% average PASS. The largest observed counting-to-action drop is 44.5 percentage points, and additional global-click and matched local-count controls show that coordinate grounding explains only part of the loss.

For comparable reporting, please include:

1. strict PASS on each of the five visual subsets;
2. macro-average PASS across the five original subsets;
3. template-level PASS for the five coupled tasks;
4. grammar-valid output rate (VALID);
5. model version and inference configuration;
6. prompt order and any image resizing or compression.

Do not treat the test split as a prompt-development set.

Intended use

ROSE is intended for:

• evaluating fine-grained visual comparison;
• measuring the gap between counting-oriented readouts and coordinate-level action;
• diagnosing region binding and context-conditioned selection;
• studying abstention on zero-target cases;
• evaluating exact symbolic execution in multimodal models;
• comparing local, API-based, and agentic multimodal systems under a shared protocol.

ROSE is a diagnostic benchmark and is not intended as a comprehensive simulation of open-world visual interaction.

Limitations

ROSE uses controlled grid-structured scenes and a fixed symbolic action space. This design enables exact measurement of perception, region selection, and execution, but does not cover the full complexity of natural images, free-form language, long-horizon planning, dynamic interaction, or environment feedback.

The current release covers five visual sources and five task templates. Model scores depend on the evaluated model version, prompt format, image preprocessing, and inference configuration. Human results in the paper are used primarily as a solvability reference rather than as a population-level estimate of human performance.

Data and licensing note

ROSE combines benchmark annotations and generated task metadata with visual assets drawn from multiple source types, including curated glyph renderings, public emoji renderings, and web-collected pixel-art assets. Dataset users are responsible for complying with the dataset license and any source-specific attribution or redistribution terms associated with the released visual assets.

Citation

Please cite ROSE if you use the dataset, evaluator, or derived analyses.

@misc{wang2026rose,
  title={ROSE: Benchmarking the Perception-to-Action Gap in Multimodal Models},
  author={Yihao Wang and Zijian He and Jie Ren and Keze Wang},
  year={2026},
  eprint={2606.19965},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.19965}
}

Version history

v0.1

• five fine-grained visual subsets;
• 1,512 scenes;
• 3,024 rendered images;
• 7,560 coupled task instances;
• official scene-level dev and test splits;
• strict COUNT / CLICK / SUBMIT protocol.

Contact

For questions about the dataset or paper, please contact the authors through the project page or the corresponding author, Keze Wang (kezewang@gmail.com).