SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2401.04259 · 2024

MARG: Multi-Agent Review Generation for Scientific Papers

arXiv · PDF · Open in the Atlas

Code that ran

7 of 10 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 0 have no contract verdict; 7 failures of our instrument; 3 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/2401.04259.json · badge: /api/ran/2401.04259.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
allenai/marg-reviewer canonical (classifier label)
classifier basis: census
7 ran · 3 no recorded run
FunctionStatusWhere it lives
basic_token_align Ran
ran; our contract draft was wrong, not the code
allenai/marg-reviewer/review_worker/aries/util/edit.py
code served (permissive licence) · get_code("5ebf094868fd0444")
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 tuple. 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 collections
import itertools

# Test basic_token_align with simple string sequences
# The function expects sequences of tokens (strings), where each token is a sequence of characters

# Create two token sequences where seq2 contains all characters of seq1
# but potentially with different tokenization
seq1 = ["hello", "world"]  # 2 tokens
seq2 = ["hel", "lo", "wor", "ld"]  # 4 tokens, same characters but different boundaries

# seq2_ignored_ids can be None or an iterable of token indices to ignore
seq2_ignored_ids = None

ARGS = {
    "seq1": seq1,
    "seq2": seq2,
    "seq2_ignored_ids": seq2_ignored_ids,
}
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 tuple — a tuple output has no array shape of its own; none is 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
colorify Ran
ran; our contract draft was wrong, not the code
allenai/marg-reviewer/review_worker/aries/util/color.py
code served (permissive licence) · get_code("6164eefcce2fb3e1")
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

COLOR_TABLE = {'black': {'hex': '3e424d', 'ansi': '30'}, 'red': {'hex': 'e75c58', 'ansi': '31'}, 'green': {'hex': '00a050', 'ansi': '32'}, 'yellow': {'hex': 'ddbb33', 'ansi': '33'}, 'blue': {'hex': '2090ff', 'ansi': '34'}, 'magenta': {'hex': 'd060c0', 'ansi': '35'}, 'cyan': {'hex': '60c7c7', 'ansi': '36'}, 'white': {'hex': 'c0c0b0', 'ansi': '37'}, 'strong-black': {'hex': '303030', 'ansi': '90'}, 'strong-red': {'hex': 'b03030', 'ansi': '91'}, 'strong-green': {'hex': '007030', 'ansi': '92'}, 'strong-yellow': {'hex': 'b08010', 'ansi': '93'}, 'strong-blue': {'hex': '0070dd', 'ansi': '94'}, 'strong-magenta': {'hex': 'a03090', 'ansi': '95'}, 'strong-cyan': {'hex': '209090', 'ansi': '96'}, 'strong-white': {'hex': 'a0a0b0', 'ansi': '97'}}

def colorify(s: str, color: str, bold: bool=False, form='html', tag_side='both'):
    """if tag_side is 'left', only the left tag is added.  If tag_side irght
    'right', only the right tag is added.  This is useful if, for example,
    a list of tokens needs to be colored without joining the tokens.  Raises an
    error if this is not possible for the given form."""
    if color is None or form == 'none':
        return s
    m = re.match('#(?P<hexcode>[0-9a-fA-F]{6})', color)
    valid_ansi = False
    if not m:
        if color in COLOR_TABLE:
            valid_ansi = True
            hex_color = COLOR_TABLE[color]['hex']
        else:
            raise ValueError('Invalid color {}'.format(color))
    else:
        hex_color = m.group('hexcode')
    left_tag, right_tag = ('', '')
    if form == 'html':
        bold_code = 'font-weight: bold;' if bold else ''
        left_tag = '<span style="color: #{code};{boldness}">'.format(code=hex_color, boldness=bold_code)
        right_tag = '</span>'
    elif form == 'ansi' and valid_ansi:
        bold_code = '1' if bold else '0'
        left_tag = '\x1b[{boldness};{code}m'.format(code=COLOR_TABLE[color]['ansi'], boldness=bold_code)
        right_tag = '\x1b[0m'
    else:
        raise ValueError('Invalid format {}'.format(form))
    if tag_side == 'left':
        return left_tag + s
    elif tag_side == 'right':
        return s + right_tag
    elif tag_side == 'both':
        return left_tag + s + right_tag
    raise ValueError('Invalid tag_side {}'.format(tag_side))

ARGS = {
    's': 'Hello World',
    'color': 'red',
    'bold': True,
    'form': 'html',
    'tag_side': 'both'
}
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
downsample_recs Ran
ran; our contract draft was wrong, not the code
allenai/marg-reviewer/review_worker/aries/util/data.py
code served (permissive licence) · get_code("28f1982db3388dba")
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 list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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 numpy as np

# Create sample records - can be any objects
recs = [
    {"id": 1, "value": "a"},
    {"id": 2, "value": "b"},
    {"id": 3, "value": "c"},
    {"id": 4, "value": "d"},
    {"id": 5, "value": "e"},
]

# Create downsample config with keep_ratio
downsample_config = {
    'keep_ratio': 0.6,
    'seed': 42
}

ARGS = {
    'recs': recs,
    'downsample_config': downsample_config
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
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
index_by Ran
ran; our contract draft was wrong, not the code
allenai/marg-reviewer/review_worker/aries/util/data.py
code served (permissive licence) · get_code("00448a446842cc38")
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 torch
import numpy as np
from typing import Union, Callable, Dict, Iterable, Iterator, List

torch.manual_seed(0)

# Test case 1: list of dicts with string key
list_of_dicts = [
    {'id': 'a', 'value': 10},
    {'id': 'b', 'value': 20},
    {'id': 'c', 'value': 30},
]

# Test case 2: list of objects with callable key
class Record:
    def __init__(self, name, score):
        self.name = name
        self.score = score

records = [
    Record('alice', 100),
    Record('bob', 95),
    Record('charlie', 100),
]

# Test case 3: iterator (generator)
def record_generator():
    yield {'id': 1, 'tag': 'x'}
    yield {'id': 2, 'tag': 'y'}
    yield {'id': 3, 'tag': 'x'}

# Test case 4: list of tuples
list_of_tuples = [
    (1, 'apple'),
    (2, 'banana'),
    (1, 'apricot'),
]

ARGS = {
    'lst': list_of_dicts,
    'key': 'id',
    'one_to_one': False,
}
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 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
openc Ran
ran; our contract draft was wrong, not the code
allenai/marg-reviewer/review_worker/aries/util/data.py
code served (permissive licence) · get_code("7142219531463a02")
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 TextIOWrapper. 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 gzip
import lzma
import tempfile
import os

# Create a temporary file with .gz extension for testing
temp_dir = tempfile.mkdtemp()
test_file_gz = os.path.join(temp_dir, "test.txt.gz")
test_file_xz = os.path.join(temp_dir, "test.txt.xz")
test_file_plain = os.path.join(temp_dir, "test.txt")

# Create a simple test file
with open(test_file_plain, 'w') as f:
    f.write("test content")

ARGS = {
    'fname': test_file_plain,
    'mode': 'rt',
    'compression': 'auto',
}
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 TextIOWrapper — a TextIOWrapper output has no array shape of its own; none is 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
raw_table2str Ran
ran; our contract draft was wrong, not the code
allenai/marg-reviewer/review_worker/aries/util/logging.py
code served (permissive licence) · get_code("ab60b777a46febd1")
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.

# Analyzing the function raw_table2str:
# - grid: a 2D list of strings (rows of cells)
# - colsep: string separator between columns
# - rowend: string to append at end of each row
# - header_row_sep: string separator after header row

# The function formats a table as a string
# grid is accessed as grid[y][x] where y is row index, x is col index
# Each cell should be a string

ARGS = {
    'grid': [
        ['Name', 'Age', 'City'],
        ['Alice', '30', 'NYC'],
        ['Bob', '25', 'LA']
    ],
    'colsep': ' | ',
    'rowend': '',
    'header_row_sep': '-'
}
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
table2str Ran
ran; our contract draft was wrong, not the code
allenai/marg-reviewer/review_worker/aries/util/logging.py
code served (permissive licence) · get_code("74be48c9b4dc9c99")
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

torch.manual_seed(0)

# Analyze the function:
# - grid: a list of lists of strings (2D grid)
# - format_fn: a callable that converts cell values to strings
# - col_names: list-like of column names
# - row_names: list-like of row names
# - colsep: string separator between columns
# - rowend: string to append at end of each row
# - header_row_sep: string separator after header row

# Create a small grid of numeric values (3 rows x 4 columns)
grid = [
    [1, 2, 3, 4],
    [5, 6, 7, 8],
    [9, 10, 11, 12]
]

# Column names
col_names = ['A', 'B', 'C', 'D']

# Row names
row_names = ['Row1', 'Row2', 'Row3']

# Format function - converts to string
def format_fn(x):
    return str(x)

# Column separator
colsep = ' | '

# Row end marker
rowend = ''

# Header row separator
header_row_sep = '-'

ARGS = {
    'grid': grid,
    'format_fn': format_fn,
    'col_names': col_names,
    'row_names': row_names,
    'colsep': colsep,
    'rowend': rowend,
    'header_row_sep': header_row_sep
}
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
colorprint Not yet run
unverified (harvested, no recorded run)
allenai/marg-reviewer/review_worker/aries/util/color.py
code served (permissive licence) · get_code("6b96792de473f699")
How it ran

no invocation recorded.

levenshtein_distance Not yet run
unverified (harvested, no recorded run)
allenai/marg-reviewer/review_worker/aries/util/edit.py
code served (permissive licence) · get_code("519a75e4f9f6bd58")
How it ran

no invocation recorded.

make_word_diff Not yet run
unverified (harvested, no recorded run)
allenai/marg-reviewer/review_worker/aries/util/edit.py
code served (permissive licence) · get_code("843af5ddd4587178")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 8 are also present in other papers whose code we hold (8 with at least one paper dated earlier, 0 the same month, 0 later) and 2 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.

FunctionAlso present inEarliest-datedOthers named
basic_token_align 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews
colorify 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews
colorprint 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews
downsample_recs 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews
index_by 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews
levenshtein_distance 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews
make_word_diff 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews
openc 1 other paper
1 dated earlier
2306.12587 (2023-06)
ARIES: A Corpus of Scientific Paper Edits Made in Response to Peer Reviews

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

We study the ability of LLMs to generate feedback for scientific papers and develop MARG, a feedback generation approach using multiple LLM instances that engage in internal discussion. By distributing paper text across agents, MARG can consume the full text of papers beyond the input length limitations of the base LLM, and by specializing agents and incorporating sub-tasks tailored to different comment types (experiments, clarity, impact) it improves the helpfulness and specificity of feedback. In a user study, baseline methods using GPT-4 were rated as producing generic or very generic comments more than half the time, and only 1.7 comments per paper were rated as good overall in the best baseline. Our system substantially improves the ability of GPT-4 to generate specific and helpful feedback, reducing the rate of generic comments from 60% to 29% and generating 3.7 good comments per paper (a 2.2x improvement).

For agents

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

get_harvested_code_for_paper("2401.04259")
get_code_for_paper("2401.04259")
have("2401.04259")

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

curl https://syntology.ai/api/ran/2401.04259.json

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

[![Syntology run record](https://syntology.ai/api/ran/2401.04259.svg)](https://syntology.ai/paper/2401.04259)

Connect an agent — have() is free.