SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.01736 · 2026

Argument Collapse: LLMs Flatten Long-Form Public Debate

Yapei Chang, Yekyung Kim, Chau Minh, Pham Mohit Iyyer

arXiv · PDF

Code that ran

16 of 18 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; 2 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 14 failures of our instrument; 2 no recorded run), measured 2026-11-22; 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/2606.01736.json · badge: /api/ran/2606.01736.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
mungg/argument_collapse canonical (classifier label)
classifier basis: census
16 ran · 2 no recorded run
FunctionStatusWhere it lives
build_adj Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/metric.py
code served (permissive licence) · get_code("d5fc6ffebec7d8b3")
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 Dict, Set, Tuple, FrozenSet

# Construct valid arguments for build_adj function

# sub_to_essay: maps sub-argument IDs to their essay IDs
sub_to_essay = {
    "sub_arg_1": "essay_A",
    "sub_arg_2": "essay_A",
    "sub_arg_3": "essay_B",
    "sub_arg_4": "essay_B",
    "sub_arg_5": "essay_C",
}

# pair_rel: maps pairs of sub-arguments to their relation type
pair_rel = {
    ("sub_arg_1", "sub_arg_3"): "entailment",
    ("sub_arg_1", "sub_arg_4"): "neutral",
    ("sub_arg_2", "sub_arg_3"): "entailment",
    ("sub_arg_3", "sub_arg_5"): "contradiction",
    ("sub_arg_1", "sub_arg_2"): "entailment",  # Same essay, should be skipped
    ("sub_arg_2", "sub_arg_5"): "neutral",
}

# relation_set: frozenset of relations we care about
relation_set = frozenset({"entailment", "contradiction"})

ARGS = {
    "sub_to_essay": sub_to_essay,
    "pair_rel": pair_rel,
    "relation_set": relation_set,
}
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
detect_layout Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/data.py
code served (permissive licence) · get_code("949f421fdd434463")
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 os
from pathlib import Path
import tempfile
import shutil

# Create a temporary directory structure for testing
temp_dir = tempfile.mkdtemp()

# Test case 1: LAYOUT_AGGREGATE - create nyt directory
nyt_dir = Path(temp_dir) / 'nyt'
nyt_dir.mkdir(exist_ok=True)

# Alternatively, we could test other layouts by creating different structures
# But for simplicity, we'll use the aggregate layout

data_root_arg = temp_dir

ARGS = {
    'data_root': data_root_arg
}
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
essay_unique Ran
ran, honoured the contract we drafted
mungg/argument_collapse/src/argument_collapse/metric.py
code served (permissive licence) · get_code("1a3e4ee14111fb31")
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 arguments for essay_unique function
essay_sub_ids = ["sub_1", "sub_2", "sub_3"]
other_sub_ids = ["sub_4", "sub_5", "sub_2"]

pair_rel = {
    ("sub_1", "sub_4"): "similar",
    ("sub_2", "sub_5"): "related",
    ("sub_3", "sub_2"): "different",
    ("sub_1", "sub_2"): "similar",
}

relation_set = frozenset(["similar", "related"])

ARGS = {
    "essay_sub_ids": essay_sub_ids,
    "other_sub_ids": other_sub_ids,
    "pair_rel": pair_rel,
    "relation_set": relation_set,
}
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
extract_json Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/pair_comparison_main_arg.py
code served (permissive licence) · get_code("b933cab935ecc2aa")
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 re
from typing import Any, Iterable

# Test cases for extract_json function
# The function takes a single parameter: text (str)

ARGS = {
    "text": '{"key": "value", "number": 42}'
}
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_json Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/stance.py
code served (permissive licence) · get_code("ec41fe411ec4206b")
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

# The function extract_json takes a single parameter: text (str)
# It processes the text to extract and parse JSON from it.
# No models, tensors, or complex objects are needed.

ARGS = {
    "text": '```json\n{"key": "value", "number": 42}\n```'
}
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_json Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/structure.py
code served (permissive licence) · get_code("ba2ef2aa84eacd6c")
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 re
from typing import Any

torch_imported = False
try:
    import torch
    torch_imported = True
except ImportError:
    pass

# Analyze the function: extract_json takes one parameter: text (str)
# The function processes a string to extract JSON from it
# It handles markdown code blocks, finds opening braces, tracks depth and strings

# Create various test cases for the text parameter
ARGS = {
    "text": '{"key": "value", "nested": {"inner": 123}}'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
jsonable Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/inference/types.py
code served (permissive licence) · get_code("cba80378bea200a8")
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 pathlib import Path
from collections.abc import Mapping

torch.manual_seed(0)

# Create a simple class with model_dump method
class PydanticLike:
    def model_dump(self):
        return {"field1": "value1", "field2": 42}

# Create a simple class with to_dict method
class DictConvertible:
    def to_dict(self):
        return {"key1": "val1", "key2": 3.14}

# Create a complex nested structure to test the recursive behavior
complex_value = {
    "none_field": None,
    "str_field": "hello",
    "int_field": 42,
    "float_field": 3.14,
    "bool_field": True,
    "path_field": Path("/home/user/data"),
    "list_field": [1, "two", 3.0, True, None],
    "tuple_field": (1, 2, 3),
    "set_field": {1, 2, 3},
    "nested_dict": {
        "inner": "value",
        "inner_path": Path("/tmp/file.txt"),
        "inner_list": [1, 2, 3]
    },
    "pydantic_obj": PydanticLike(),
    "dict_obj": DictConvertible(),
    "custom_obj": "will be converted to str"
}

ARGS = {
    "value": complex_value
}
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
normalize Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/pair_comparison_main_arg.py
code served (permissive licence) · get_code("ab25f2520bea63b4")
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

torch.manual_seed(0)

# Analyze the function:
# - normalize() takes one parameter: payload (dict[str, Any])
# - It accesses payload.get('relation', ''), payload.get('rationale', ''), 
#   payload.get('reason', ''), payload.get('explanation', '')
# - It converts these to strings, strips whitespace, and lowercases 'relation'
# - It checks if relation is in VALID_RELATIONS
# - It returns a dict with 'relation' and 'rationale' keys

ARGS = {
    'payload': {
        'relation': 'equivalent',
        'rationale': 'These two items are semantically the same.',
        'reason': 'backup reason',
        'explanation': 'backup explanation'
    }
}
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
normalize_annotation Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/structure.py
code served (permissive licence) · get_code("ab1d6a9928b9ae6f")
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.

ARGS = {
    'payload': {
        'annotations': [
            {'paragraph_index': 0, 'labels': ['label1', 'label2'], 'rationale': 'This is a rationale'},
            {'paragraph_index': 1, 'labels': 'single_label', 'rationale': 'Another rationale'},
            {'paragraph_index': 2, 'labels': ['label3'], 'rationale': ''}
        ]
    },
    'layer': 'argument',
    'n_paragraphs': 3
}
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 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
normalize_venue Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/data.py
code served (permissive licence) · get_code("91b783517dd37d3d")
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

ARGS = {
    'venue': 'nyt'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
parse_frontmatter_and_body Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/data.py
code served (permissive licence) · get_code("4b3d96b8a4765eb1")
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, 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.

from pathlib import Path
import tempfile
import os

# Create a temporary directory and file with frontmatter
temp_dir = tempfile.mkdtemp()
test_file_path = Path(temp_dir) / "test.md"

# Write a markdown file with YAML frontmatter
markdown_content = """---
title: Test Document
author: John Doe
date: 2024-01-15
---
This is the body of the markdown file.

It can have multiple lines and paragraphs.
"""

test_file_path.write_text(markdown_content)

ARGS = {
    "path": test_file_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 tuple, ndim 1, dtype family float — a tuple 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
split_paragraphs Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/structure.py
code served (permissive licence) · get_code("57f9aa481a38511d")
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 re

def split_paragraphs(body: str) -> list[str]:
    return [p.strip() for p in re.split('\\n\\s*\\n+', body.strip()) if p.strip()]

# Construct valid arguments
body = """This is the first paragraph.
It has multiple lines.

This is the second paragraph.
Also multiple lines here.

And a third paragraph
with some content."""

ARGS = {
    "body": body
}
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 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
user_prompt Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/pair_comparison_main_arg.py
code served (permissive licence) · get_code("cf84555b316c3ce4")
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

torch.manual_seed(0)

ARGS = {
    'question': 'Should artificial intelligence be regulated by governments?',
    'main_i': 'AI regulation is necessary to prevent misuse and ensure safety standards across industries.',
    'main_j': 'Heavy regulation stifles innovation and competitive advantage in the AI sector.',
    'context_label': 'Debate question'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
user_prompt Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/pair_comparison_sub_arg.py
code served (permissive licence) · get_code("2b92e56e678bfc1e")
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

torch.manual_seed(0)

ARGS = {
    'question': 'Should artificial intelligence be regulated by governments?',
    'main_i': 'AI regulation is necessary to protect public safety and prevent misuse.',
    'sub_i': 'Unregulated AI systems could be used for harmful purposes without oversight.',
    'main_j': 'AI should develop freely without government interference to maximize innovation.',
    'sub_j': 'Excessive regulation stifles technological progress and economic growth.',
    'context_label': 'Debate question'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
user_prompt Ran
ran; our contract draft was wrong, not the code
mungg/argument_collapse/src/argument_collapse/annotate/toulmin.py
code served (permissive licence) · get_code("c5164adfd99fcd74")
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)

ARGS = {
    'context': 'Is artificial intelligence beneficial to society?',
    'body': 'Artificial intelligence has transformative potential across multiple domains including healthcare, education, and scientific research. However, careful regulation is needed to ensure ethical deployment.',
    'context_label': 'Debate question'
}
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
within_unique Ran
ran, honoured the contract we drafted
mungg/argument_collapse/src/argument_collapse/metric.py
code served (permissive licence) · get_code("8ce24061cc5dd623")
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
from math import comb

# Set seed for reproducibility
torch.manual_seed(0)

# Construct arguments for within_unique function
# The function needs:
# 1. sub_to_essay: dict mapping sub-argument IDs to essay IDs
# 2. pair_rel: dict mapping (sub_arg1, sub_arg2) tuples to relation strings
# 3. m: int for sample size
# 4. relation_set: frozenset or set of relation strings to consider

# Build a small example with 2 essays and 4 sub-arguments
sub_to_essay = {
    "sub_1": "essay_A",
    "sub_2": "essay_A",
    "sub_3": "essay_B",
    "sub_4": "essay_B",
}

# Build pair_rel with relations between sub-arguments
# Key format: (sub_arg_id_1, sub_arg_id_2)
# Only cross-essay pairs will be considered (same essay pairs are filtered)
pair_rel = {
    ("sub_1", "sub_3"): "related",      # essay_A to essay_B
    ("sub_1", "sub_4"): "unrelated",    # essay_A to essay_B
    ("sub_2", "sub_3"): "related",      # essay_A to essay_B
    ("sub_2", "sub_4"): "unrelated",    # essay_A to essay_B
    ("sub_1", "sub_2"): "related",      # same essay (will be filtered)
    ("sub_3", "sub_4"): "related",      # same essay (will be filtered)
}

# Sample size m for the within_unique calculation
m = 2

# Relation set - only "related" relations will create edges in adjacency
relation_set = frozenset(["related"])

ARGS = {
    "sub_to_essay": sub_to_essay,
    "pair_rel": pair_rel,
    "m": m,
    "relation_set": relation_set,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
build_clusters Not yet run
unverified (harvested, no recorded run)
mungg/argument_collapse/src/argument_collapse/cluster.py
code served (permissive licence) · get_code("7d061e7003a2a8e6")
How it ran

no invocation recorded.

select_llm_representatives Not yet run
unverified (harvested, no recorded run)
mungg/argument_collapse/src/argument_collapse/cluster.py
code served (permissive licence) · get_code("c2b600030962f3b0")
How it ran

no invocation recorded.

Also present in other papers' code

Of 18 samples lifted from this paper's linked repositories, 18 carry a function identity and 0 do not. Of the 16 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 16 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

As LLMs are increasingly used to draft publicfacing arguments, they may flatten public debate by repeatedly introducing the same polished, plausible arguments. We study argument collapse, the tendency of essays generated by different LLMs to converge to a smaller set of main arguments, sub-arguments, and paragraph-level structures. We compare 1,039 human responses from 195 New York Times (NYT) debates, 448 human responses from 61 longer-form Boston Review (BR) forums, and 23,384 LLM-generated essays. In the NYT corpus, 65.3% of human main arguments are unique within a debate, compared to 3.4% of LLM main arguments. Asking LLMs to generate diverse answers adds variation, but a typical model recovers only about half of the distinct human main arguments, with much of the added variation falling outside the observed human argument space. Collapse also appears in sub-arguments, where among essays with the same main argument, 41.0% of human subarguments are unique versus 9.1% from LLM responses. Qualitatively, LLMs often reuse generalized and hedged sub-arguments, while humans prefer more concrete and topic-specific ones. Structure-wise, LLM-generated essays tend to follow a more fixed arc, often opening with a direct claim and moving quickly toward proposals. The same patterns hold in longer BR essays, suggesting that argument collapse extends beyond short-form responses. github.com/mungg/argument_collapse

For agents

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

get_harvested_code_for_paper("2606.01736")
get_code_for_paper("2606.01736")
have("2606.01736")

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

curl https://syntology.ai/api/ran/2606.01736.json

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

[![Syntology run record](https://syntology.ai/api/ran/2606.01736.svg)](https://syntology.ai/paper/2606.01736)

Connect an agent — have() is free.