SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2409.02813 · 2024

MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark

arXiv · PDF · Open in the Atlas

Code that ran

11 of 12 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 11 have no contract verdict; 0 failures of our instrument; 1 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2409.02813.json · badge: /api/ran/2409.02813.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
opendatalab/pm4bench community (classifier label)
classifier basis not recorded on the link
11 ran · 1 no recorded run
FunctionStatusWhere it lives
aggregate_scores Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/miqa.py
code served (permissive licence) · get_code("cc5c4cade54886a7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Construct valid records for the aggregate_scores function
records = [
    {
        'scores': {
            'Creativity': 8,
            'Richness': 7,
            'Visual Perception': 9,
            'Logical Coherence': 8,
            'Answer Accuracy': 9,
            'Image Relationship Understanding': 7
        }
    },
    {
        'scores': {
            'Creativity': 6,
            'Richness': 8,
            'Visual Perception': 7,
            'Logical Coherence': 9,
            'Answer Accuracy': 8,
            'Image Relationship Understanding': 8
        }
    },
    {
        'scores': {
            'Creativity': 9,
            'Richness': 6,
            'Visual Perception': 8,
            'Logical Coherence': 7,
            'Answer Accuracy': 7,
            'Image Relationship Understanding': 9
        }
    }
]

ARGS = {
    'records': records
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compare_gt Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/mgui/render.py
code served (permissive licence) · get_code("40019c900986263c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
from pathlib import Path

# Create temporary directories for candidate and reference roots
candidate_root = Path(tempfile.mkdtemp())
reference_root = Path(tempfile.mkdtemp())

# Create sample JSON files for comparison
candidate_file = candidate_root / "test.json"
reference_file = reference_root / "test.json"

candidate_data = {
    "template": "template_1",
    "template_title": "Title 1",
    "lang": "en",
    "image": "image.png",
    "width": 800,
    "height": 600,
    "all_qid_boxes": {
        "Q1": {
            "x": 10.0,
            "y": 20.0,
            "width": 50.0,
            "height": 30.0,
            "text": "<built-in method copy of dict object at 0x12345678>"
        },
        "Q2": {
            "x": 100.0,
            "y": 150.0,
            "width": 60.0,
            "height": 40.0,
            "text": "Sample text"
        }
    },
    "questions": [
        {
            "qid_in_template": "Q1",
            "difficulty": "easy",
            "target_qid": "Q1",
            "question": "What is this?",
            "target_text_sample": "Text sample 1"
        },
        {
            "qid_in_template": "Q2",
            "difficulty": "hard",
            "target_qid": "Q2",
            "question": "What is that?",
            "target_text_sample": "Text sample 2"
        }
    ]
}

reference_data = {
    "template": "template_1",
    "template_title": "Title 1",
    "lang": "en",
    "image": "image.png",
    "width": 800,
    "height": 600,
    "all_qid_boxes": {
        "Q1": {
            "x": 10.5,
            "y": 20.1,
            "width": 50.0,
            "height": 30.0,
            "text": "<built-in method copy of dict object at 0x87654321>"
        },
        "Q2": {
            "x": 100.0,
            "y": 150.0,
            "width": 60.0,
            "height": 40.0,
            "text": "Sample text"
        }
    },
    "questions": [
        {
            "qid_in_template": "Q1",
            "difficulty": "easy",
            "target_qid": "Q1",
            "question": "What is this?",
            "target_text_sample": "Text sample 1"
        },
        {
            "qid_in_template": "Q2",
            "difficulty": "hard",
            "target_qid": "Q2",
            "question": "What is that?",
            "target_text_sample": "Text sample 2"
        }
    ]
}

# Write JSON files
with candidate_file.open('w', encoding='utf-8') as f:
    json.dump(candidate_data, f)

with reference_file.open('w', encoding='utf-8') as f:
    json.dump(reference_data, f)

ARGS = {
    "candidate_root": candidate_root,
    "reference_root": reference_root,
    "tolerance": 0.2,
    "compare_geometry": True
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
extract_answer_letter Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/metrics.py
code served (permissive licence) · get_code("e929aabca68bb917")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def extract_answer_letter(value: str) -> str | None:
    text = re.sub('<think>.*?</think>', '', value or '', flags=re.DOTALL)
    explicit = re.findall('(?:answer|option)\\s*[::]?\\s*\\(?([A-J])\\)?', text, re.IGNORECASE)
    if explicit:
        return explicit[-1].upper()
    boxed = re.findall('\\\\boxed\\{\\s*([A-J])\\s*\\}', text, re.IGNORECASE)
    if boxed:
        return boxed[-1].upper()
    standalone = re.findall('(?<![A-Za-z])([A-J])(?![A-Za-z])', text, re.IGNORECASE)
    return standalone[-1].upper() if standalone else None

# Test case 1: explicit answer format
test_value_1 = "The answer is: B"

# Test case 2: boxed format
test_value_2 = "The final answer is \\boxed{C}"

# Test case 3: with think tags
test_value_3 = "<think>working through this</think>Answer: D"

# Test case 4: standalone letter
test_value_4 = "The result is E"

# Test case 5: None/empty string
test_value_5 = None

ARGS = {
    "value": test_value_1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
extract_scores Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/miqa.py
code served (permissive licence) · get_code("6ec5c1760b10f4d4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import ast
import re

DIMENSIONS = ('Creativity', 'Richness', 'Visual Perception', 'Logical Coherence', 'Answer Accuracy', 'Image Relationship Understanding')

def _normalize_scores(value: dict, ) -> dict | None:
    canonical = {dimension.lower(): dimension for dimension in DIMENSIONS + ('Overall Score',)}
    result = {}
    for key, score in value.items():
        name = canonical.get(str(key).strip().lower())
        if not name:
            continue
        try:
            number = int(score)
        except (TypeError, ValueError):
            continue
        if 1 <= number <= 10:
            result[name] = number
    return result if all((dimension in result for dimension in DIMENSIONS)) else None

def extract_scores(text: str) -> dict | None:
    if not text:
        return None
    candidates = ['{' + match + '}' for match in re.findall('\\{\\{(.*?)\\}\\}', text, re.DOTALL)]
    candidates.extend(re.findall('\\{[^{}]*\\}', text, re.DOTALL))
    for candidate in reversed(candidates):
        try:
            value = ast.literal_eval(candidate)
        except (SyntaxError, ValueError):
            continue
        if isinstance(value, dict):
            normalized = _normalize_scores(value)
            if normalized:
                return normalized
    pairs = re.findall('[\'\\"]?([A-Za-z][A-Za-z ]+?)[\'\\"]?\\s*[::]\\s*(\\d{1,2})', text)
    return _normalize_scores(dict(pairs))


# Test case 1: Text with dictionary in double braces
text1 = """
{{
    'Creativity': 8,
    'Richness': 7,
    'Visual Perception': 9,
    'Logical Coherence': 8,
    'Answer Accuracy': 9,
    'Image Relationship Understanding': 8
}}
"""

# Test case 2: Text with key-value pairs in colon format
text2 = """
Creativity: 7
Richness: 8
Visual Perception: 6
Logical Coherence: 9
Answer Accuracy: 8
Image Relationship Understanding: 7
"""

# Test case 3: Text with dictionary in single braces
text3 = """{
    "Creativity": 5,
    "Richness": 6,
    "Visual Perception": 7,
    "Logical Coherence": 8,
    "Answer Accuracy": 9,
    "Image Relationship Understanding": 10
}"""

# Test case 4: Empty string
text4 = ""

# Test case 5: Text with Chinese colon (:)
text5 = """
Creativity:9
Richness:8
Visual Perception:7
Logical Coherence:9
Answer Accuracy:8
Image Relationship Understanding:9
"""

ARGS = {
    "text": text1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
format_reward Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/qgo/reward.py
code served (permissive licence) · get_code("c17d2901ecfbf4cb")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import math

REPETITION_THRESHOLD = 0.6
REPETITION_PENALTY_WEIGHT = 0.4
LENGTH_PENALTY_WEIGHT = 0.8
LENGTH_REWARD_WEIGHT = 0.2

def format_reward(solution: str, minimum_length: int=1000, maximum_length: int=10000) -> dict[str, float]:
    length = len(solution)
    if '[Internal Processing Omitted]' in solution:
        return {'combined': -1.2, 'length': 0.0, 'length_penalty': 0.0, 'repetition': 0.0}
    check_text = solution[-3000:] if length > 5000 else solution
    ngrams = [check_text[index:index + 4] for index in range(max(0, len(check_text) - 3))]
    diversity = len(set(ngrams)) / len(ngrams) if ngrams else 1.0
    repetition = max(0.0, 1.0 - diversity / REPETITION_THRESHOLD)
    length_reward = float(minimum_length <= length <= maximum_length)
    if length < minimum_length:
        length_penalty = 1.0 - length / minimum_length
    elif length > maximum_length:
        length_penalty = min(1.0, math.log1p((length - maximum_length) / 1200.0))
    else:
        length_penalty = 0.0
    combined = length_reward * LENGTH_REWARD_WEIGHT - length_penalty * LENGTH_PENALTY_WEIGHT - repetition * REPETITION_PENALTY_WEIGHT
    return {'combined': combined, 'length': length_reward, 'length_penalty': length_penalty, 'repetition': repetition}


# Create a solution string within the valid length range
solution_text = "This is a test solution. " * 100

ARGS = {
    'solution': solution_text,
    'minimum_length': 1000,
    'maximum_length': 10000
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_predictions Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/io.py
code served (permissive licence) · get_code("e6b5d81f514b3ebb")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import json
from pathlib import Path

# Create a temporary JSONL file with valid records
with tempfile.NamedTemporaryFile(mode='w', suffix='.jsonl', delete=False, encoding='utf-8') as f:
    temp_path = f.name
    # Write valid records with 'id' field
    f.write(json.dumps({'id': '1', 'prediction': 'value1'}) + '\n')
    f.write(json.dumps({'id': '2', 'prediction': 'value2'}) + '\n')
    # Write a record with 'index' field
    f.write(json.dumps({'index': 3, 'prediction': 'value3'}) + '\n')
    # Include an empty line to test line skipping
    f.write('\n')
    f.write(json.dumps({'id': '4', 'prediction': 'value4'}) + '\n')

ARGS = {
    'path': temp_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
mdur_correct Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/metrics.py
code served (permissive licence) · get_code("cd1307d13954b765")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def extract_answer_letter(value: str) -> str | None:
    text = re.sub('<think>.*?</think>', '', value or '', flags=re.DOTALL)
    explicit = re.findall('(?:answer|option)\\s*[::]?\\s*\\(?([A-J])\\)?', text, re.IGNORECASE)
    if explicit:
        return explicit[-1].upper()
    boxed = re.findall('\\\\boxed\\{\\s*([A-J])\\s*\\}', text, re.IGNORECASE)
    if boxed:
        return boxed[-1].upper()
    standalone = re.findall('(?<![A-Za-z])([A-J])(?![A-Za-z])', text, re.IGNORECASE)
    return standalone[-1].upper() if standalone else None

def mdur_correct(response: str, answer_key: str) -> bool:
    return extract_answer_letter(response) == answer_key.strip().upper()


ARGS = {
    "response": "The answer is (B). Let me think about this problem.",
    "answer_key": "b"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_historical_text Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/mgui/render.py
code served (permissive licence) · get_code("3ad8b96dec73d6f2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

HISTORICAL_METHOD_ADDRESS = re.compile('(<built-in method (?:copy|update) of dict object at )0x[0-9a-fA-F]+(>)')

def normalize_historical_text(value: str) -> str:
    """Ignore only nondeterministic addresses emitted by two legacy templates."""
    return HISTORICAL_METHOD_ADDRESS.sub('\\g<1>0x<address>\\2', value)

# Test input with a historical method address pattern
test_value = "<built-in method copy of dict object at 0x7f1a2b3c4d5e>"

ARGS = {
    "value": test_value
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_text Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/metrics.py
code served (permissive licence) · get_code("abd00a80d4ff626d")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re
import unicodedata

def normalize_text(value: str) -> str:
    return re.sub('\\s+', '', unicodedata.normalize('NFKC', value or ''))

# The function takes a single parameter: value (a string)
# It normalizes the string using NFKC normalization and removes all whitespace
# We need to provide a valid string argument

ARGS = {
    'value': '  Hello   World  \n\t  Test  '
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
record_key Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/io.py
code served (permissive licence) · get_code("cbd1256114d18a72")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# The function record_key takes a dict with optional 'id' or 'index' keys
# and returns a string representation of whichever is present (id takes priority)

# Create a valid record dict with an 'id' field
record_with_id = {
    'id': 42,
    'name': 'test_record',
    'data': [1, 2, 3]
}

ARGS = {
    'record': record_with_id
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
resolve_asset Ran
ran on a synthesized input
opendatalab/pm4bench/src/pm4bench/io.py
code served (permissive licence) · get_code("791bca2fe997c4f0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned PosixPath. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from pathlib import Path
import tempfile
import os

# Create a temporary directory to serve as the dataset root
temp_dir = tempfile.mkdtemp()
dataset_root = temp_dir

# Create a subdirectory structure for testing
subdir = os.path.join(temp_dir, "assets")
os.makedirs(subdir, exist_ok=True)

# Create a test file within the root
test_file_path = os.path.join(subdir, "test_asset.txt")
with open(test_file_path, "w") as f:
    f.write("test")

# Define relative path that stays within the root
relative_path = "assets/test_asset.txt"

ARGS = {
    "dataset_root": dataset_root,
    "relative_path": relative_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type PosixPath — a PosixPath output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PosixPath output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
render_mgui Not yet run
unverified (harvested, no recorded run)
opendatalab/pm4bench/src/pm4bench/mgui/render.py
code served (permissive licence) · get_code("9a98d066a6e55e2e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 12 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

This paper introduces MMMU-Pro, a robust version of the Massive Multi-discipline Multimodal Understanding and Reasoning (MMMU) benchmark. MMMU-Pro rigorously assesses multimodal models' true understanding and reasoning capabilities through a three-step process based on MMMU: (1) filtering out questions answerable by text-only models, (2) augmenting candidate options, and (3) introducing a vision-only input setting where questions are embedded within images. This setting challenges AI to truly "see" and "read" simultaneously, testing a fundamental human cognitive skill of seamlessly integrating visual and textual information. Results show that model performance is substantially lower on MMMU-Pro than on MMMU, ranging from 16.8% to 26.9% across models. We explore the impact of OCR prompts and Chain of Thought (CoT) reasoning, finding that OCR prompts have minimal effect while CoT generally improves performance. MMMU-Pro provides a more rigorous evaluation tool, closely mimicking real-world scenarios and offering valuable directions for future research in multimodal AI.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2409.02813")
get_code_for_paper("2409.02813")
have("2409.02813")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2409.02813.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2409.02813.svg)](https://syntology.ai/paper/2409.02813)

Connect an agent — have() is free.