SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2403.05307 · 2024

Tapilot-Crossing: Benchmarking and Evolving LLMs Towards Interactive Data Analysis Agents

arXiv · PDF · Open in the Atlas

Code that ran

16 of 16 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, 3 violated it and 0 have no contract verdict; 13 failures of our instrument; 0 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/2403.05307.json · badge: /api/ran/2403.05307.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
tapilot-crossing/tapilot_code canonical (classifier label)
classifier basis: census
16 ran · 0 no recorded run
FunctionStatusWhere it lives
extract_answer Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/eval/eval_multi_choice.py
code served (permissive licence) · get_code("8c5a9f85651f57df")
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(text_with_answer):
    if '</choice>' in text_with_answer and '\nAnswer:' not in text_with_answer:
        choice_match = re.search('\\b([A-J])\\.?\\b[^<]*</choice>', text_with_answer)
        if choice_match:
            return choice_match.group(1).upper()
    if '\nAnswer:' in text_with_answer and '</choice>' not in text_with_answer:
        answer_match = re.search('\\nAnswer: \\b([A-J])\\.?\\b', text_with_answer)
        if answer_match:
            return answer_match.group(1).upper()
    if '\nAnswer:' not in text_with_answer and '</choice>' not in text_with_answer:
        choice_match = re.search('([A-J])\\.?\\b[^<]*', text_with_answer)
        if choice_match:
            return choice_match.group(1).upper()
    return ''

ARGS = {
    'text_with_answer': 'The answer is C.</choice>'
}
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_code Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_AIR.py
code served (permissive licence) · get_code("c805463b7efcc6de")
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
import tempfile

# Create a temporary directory and files for the function to read
temp_dir = tempfile.mkdtemp()

# Create a prompt_curr.txt file with the expected format
prompt_content = """Some initial prompt content
[YOU (AI assistant)]
Here is the current state
'''
# Initial code header
x = 1
"""

with open(os.path.join(temp_dir, 'prompt_curr.txt'), 'w') as f:
    f.write(prompt_content)

# Create sample LLM response with Step 2
llm_response = """Some analysis here
# Step 2:
import numpy as np
import torch

def my_function():
    arr = np.array([1, 2, 3])
    return arr

result = my_function()

Final comment"""

ARGS = {
    'root': temp_dir,
    'llm_resp': llm_response
}
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_code Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_base.py
code served (permissive licence) · get_code("0d2c5937f51ac5ab")
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:
# - llm_resp: a string that contains either 'import ' or code with triple quotes
# - code_head: a string that serves as a prefix for code
# The function processes strings and returns a processed string

# Create test arguments
llm_resp_with_import = """Some text before
import numpy as np
import torch
def my_function():
    pass
"""

llm_resp_without_import = """def helper():
    x = 1
    return x
'''"""

code_head_value = "import sys\n"

# Test case 1: with 'import ' in response
ARGS = {
    "llm_resp": llm_resp_with_import,
    "code_head": code_head_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 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_python_code Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_COT.py
code served (permissive licence) · get_code("7289a9d629da016d")
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:
# - llm_resp: a string containing LLM response with code markers
# - code_head: a string prefix to prepend to code
# - root: a string (used only in error printing)
# The function parses llm_resp to extract Python code blocks

# Create test cases for different branches of the function
ARGS = {
    "llm_resp": "Some preamble\n---BEGIN CODE\nimport torch\nimport numpy as np\n\ndef process():\n    pass\n```",
    "code_head": "# Header\n",
    "root": "/path/to/root"
}
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
format_code Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_AIR.py
code served (permissive licence) · get_code("4b184cec1d20b4b1")
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

def format_code(code_all):
    for line in code_all.split('\n'):
        if 'plt.show(' in line or 'show_plots(' in line:
            code_all = code_all.replace(line, '')
        if 'assert' in line:
            code_all = code_all.replace(line, '')
    return code_all

# Construct valid arguments for format_code
code_sample = """
import matplotlib.pyplot as plt
def my_function():
    x = 1
    plt.show()
    assert x == 1
    print('hello')
    show_plots(x)
"""

ARGS = {
    'code_all': code_sample
}
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
get_llm_response Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/methods/utils.py
code served (permissive licence) · get_code("3373c6a0b725a712")
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 openai

# Mock the openai module since we can't make real API calls
class MockChatCompletion:
    @staticmethod
    def create(engine, messages, temperature, max_tokens, top_p, frequency_penalty, presence_penalty, timeout, stop):
        return {
            'choices': [{'message': {'content': 'Mock response'}}],
            'usage': {'prompt_tokens': 10, 'completion_tokens': 20}
        }

openai.ChatCompletion = MockChatCompletion

ARGS = {
    'messages': [
        {'role': 'system', 'content': 'You are a helpful assistant.'},
        {'role': 'user', 'content': 'What is 2+2?'}
    ],
    'engine': 'gpt-4-32k',
    'temperature': 0,
    'max_tokens': 2000,
    'top_p': 1,
    'frequency_penalty': 0,
    'presence_penalty': 0,
    'timeout': 10,
    'stop': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
openai — 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
is_float Ran
ran, violated the contract we drafted
tapilot-crossing/tapilot_code/eval/eval_code_gen.py
code served (permissive licence) · get_code("8fc4a6d9995f1727")
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 torch

def is_float(s):
    if '.' not in s:
        return s.isdecimal()
    else:
        try:
            before_decimal, after_decimal = s.split('.')
        except ValueError:
            print(s)
            print(root)
            raise AssertionError
        return before_decimal.isdecimal() and (after_decimal.isdecimal() or after_decimal == '')

ARGS = {
    's': '123.456'
}
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 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
is_folder_empty Ran
ran, violated the contract we drafted
tapilot-crossing/tapilot_code/eval/eval_multi_choice.py
code served (permissive licence) · get_code("af35b7d82ed3408a")
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 os
import tempfile
import shutil

# Create a temporary empty folder
temp_dir = tempfile.mkdtemp()

# Ensure it's empty
for item in os.listdir(temp_dir):
    item_path = os.path.join(temp_dir, item)
    if os.path.isdir(item_path):
        shutil.rmtree(item_path)
    else:
        os.remove(item_path)

ARGS = {
    "folder_path": temp_dir
}
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
is_penultimate_directory Ran
ran, violated the contract we drafted
tapilot-crossing/tapilot_code/eval/eval_code_gen.py
code served (permissive licence) · get_code("17f3b85b87f477d7")
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 os
import tempfile
import shutil

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

# Create a penultimate directory structure:
# temp_base/
#   penultimate_dir/
#     leaf_dir1/
#     leaf_dir2/

penultimate_path = os.path.join(temp_base, "penultimate_dir")
os.makedirs(penultimate_path)

# Create subdirectories with no further subdirectories
leaf_dir1 = os.path.join(penultimate_path, "leaf_dir1")
leaf_dir2 = os.path.join(penultimate_path, "leaf_dir2")
os.makedirs(leaf_dir1)
os.makedirs(leaf_dir2)

# Create some files in the leaf directories to make them non-empty
with open(os.path.join(leaf_dir1, "file1.txt"), "w") as f:
    f.write("test")
with open(os.path.join(leaf_dir2, "file2.txt"), "w") as f:
    f.write("test")

ARGS = {
    "path": penultimate_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 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
parse_chain Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/eval/eval_code_gen.py
code served (permissive licence) · get_code("5966028efcba3e7a")
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 ast

# Parse the function to understand its parameters
# The function signature is: parse_chain(node, prefix=[])
# 
# Looking at the function body:
# - node: should be an AST node (ast.Name, ast.Call, ast.Attribute, or other)
# - prefix: a list that accumulates attribute names and identifiers
#
# The function recursively processes AST nodes to extract a chain of names/attributes

# Create valid AST nodes to test the function
name_node = ast.Name(id='x', ctx=ast.Load())

# Create an Attribute node: obj.attr
attr_node = ast.Attribute(
    value=ast.Name(id='obj', ctx=ast.Load()),
    attr='method',
    ctx=ast.Load()
)

# Create a Call node: func()
call_node = ast.Call(
    func=ast.Name(id='func', ctx=ast.Load()),
    args=[],
    keywords=[]
)

# Create a complex chain: obj.attr()
chained_call = ast.Call(
    func=ast.Attribute(
        value=ast.Name(id='obj', ctx=ast.Load()),
        attr='method',
        ctx=ast.Load()
    ),
    args=[],
    keywords=[]
)

# Test with a simple Name node and default prefix
ARGS = {
    'node': name_node,
    'prefix': []
}
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
parse_reference_answer_content Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/eval/eval_multi_choice.py
code served (permissive licence) · get_code("2143bcd79a8030f5")
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 re

def parse_reference_answer_content(content):
    options_match = re.search('"options":\\s*\\[(.*?)\\]', content, re.DOTALL)
    correct_answer_match = re.search('"correct_answer":\\s*"([^"]+)"', content)
    correct_answer_match_1 = re.search('"correct_answer":\\s*([^"]+)', content)
    if correct_answer_match:
        options = []
        correct_answer = correct_answer_match.group(1).strip()
        return (options, correct_answer)
    elif correct_answer_match_1:
        options = []
        correct_answer = correct_answer_match_1.group(1).strip()
        return (options, correct_answer)
    else:
        return ([], '')

# Test case 1: content with quoted correct_answer
content_1 = '{"options": ["a", "b", "c"], "correct_answer": "b"}'

# Test case 2: content with unquoted correct_answer
content_2 = '{"options": ["x", "y"], "correct_answer": 42}'

# Test case 3: content with no correct_answer
content_3 = '{"options": ["p", "q"]}'

ARGS = {
    "content": content_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 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
prepare_code_head Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_COT.py
code served (permissive licence) · get_code("a52f99f2c2698f9d")
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
import tempfile

# Create a temporary directory with the required file
temp_dir = tempfile.mkdtemp()

# Create a sample prompt_curr.txt file with content that includes the markers
prompt_content = """Some initial content
[YOU (AI assistant)]
Some text before the triple quotes
'''
def hello():
    print("This is the code head")
    return 42
'''
"""

with open(os.path.join(temp_dir, 'prompt_curr.txt'), 'w') as f:
    f.write(prompt_content)

ARGS = {
    'root': temp_dir
}
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
prepare_code_head Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_base.py
code served (permissive licence) · get_code("bde1ccb6e2bc9262")
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.

import os
import tempfile

# Create a temporary directory and a mock file structure
temp_dir = tempfile.mkdtemp()
mock_root = os.path.join(temp_dir, "test_project_private")
os.makedirs(mock_root, exist_ok=True)

# Create a mock prompt_curr.txt file
prompt_content = """Some initial content here
[YOU (AI assistant)]
Here is the code:
'''
def helper_func():
    return 42
'''
"""

prompt_file = os.path.join(mock_root, 'prompt_curr.txt')
with open(prompt_file, 'w') as f:
    f.write(prompt_content)

# Create mock llm_response_dict with private functions
llm_response_dict_func_02 = {
    mock_root: """
    <FUNCTIONS>
    def private_helper():
        pass
    </FUNCTIONS>
    Some other content
    """
}

ARGS = {
    'root': mock_root,
    'llm_response_dict_func_02': llm_response_dict_func_02
}
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
prepare_private_funcs Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_base.py
code served (permissive licence) · get_code("9ebe5ad9b6ac2548")
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)

# Construct arguments for prepare_private_funcs
root = "test_key"

llm_response_dict_func_02 = {
    "test_key": "some prefix text <FUNCTIONS>def my_func():\n    pass</FUNCTIONS> some suffix text"
}

ARGS = {
    "root": root,
    "llm_response_dict_func_02": llm_response_dict_func_02,
}
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
process_python_output Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/methods/utils.py
code served (permissive licence) · get_code("1e9bc7efa36c49db")
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 pandas as pd
import ast
import matplotlib.pyplot as plt

# Create test stdout string with various types of outputs
stdout = """{'key1': 'value1', 'key2': 'value2'}
[1, 2, 3]
{'a': 1, 'b': 2, 'c': 3, 'd': 4, 'e': 5, 'f': 6, 'g': 7, 'h': 8, 'i': 9, 'j': 10, 'k': 11, 'l': 12, 'm': 13, 'n': 14, 'o': 15, 'p': 16}
This is a regular text line
42
True
"""

# Create a test DataFrame
df_test = pd.DataFrame({
    'col1': [1, 2, 3, 4, 5],
    'col2': ['a', 'b', 'c', 'd', 'e'],
    'col3': [10.0, 20.0, 30.0, 40.0, 50.0]
})

# Create a test Figure
fig_test = plt.figure()
ax = fig_test.add_subplot(111)
ax.plot([1, 2, 3], [1, 2, 3])

# Build stdout that includes DataFrame and Figure
stdout_with_objects = """{'key1': 'value1', 'key2': 'value2'}
[1, 2, 3]
This is a regular text line
42
"""

ARGS = {
    'stdout': stdout
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
matplotlib, pandas — 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
remove_read_csv Ran
ran; our contract draft was wrong, not the code
tapilot-crossing/tapilot_code/postprocessing/combine_code_gen_COT.py
code served (permissive licence) · get_code("12ca59656e73c371")
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

def remove_read_csv(code_add):
    for line in code_add.split('\n'):
        if 'def read_csv_file(' in line:
            continue
        if 'read_csv_file(' in line or 'pd.read_csv(' in line:
            code_add = code_add.replace(line, '')
    return code_add

# Construct valid arguments for remove_read_csv
code_add = """def some_function():
    pass

def read_csv_file(path):
    return pd.read_csv(path)

x = read_csv_file('data.csv')
y = pd.read_csv('other.csv')
z = 42
"""

ARGS = {
    'code_add': code_add
}
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

Also present in other papers' code

Of 16 samples lifted from this paper's linked repositories, 16 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

Interactive Data Analysis, the collaboration between humans and LLM agents, enables real-time data exploration for informed decision-making. The challenges and costs of collecting realistic interactive logs for data analysis hinder the quantitative evaluation of Large Language Model (LLM) agents in this task. To mitigate this issue, we introduce Tapilot-Crossing, a new benchmark to evaluate LLM agents on interactive data analysis. Tapilot-Crossing contains 1024 interactions, covering 4 practical scenarios: Normal, Action, Private, and Private Action. Notably, Tapilot-Crossing is constructed by an economical multi-agent environment, Decision Company, with few human efforts. We evaluate popular and advanced LLM agents in Tapilot-Crossing, which underscores the challenges of interactive data analysis. Furthermore, we propose Adaptive Interaction Reflection (AIR), a self-generated reflection strategy that guides LLM agents to learn from successful history. Experiments demonstrate that Air can evolve LLMs into effective interactive data analysis agents, achieving a relative performance improvement of up to 44.5%.

For agents

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

get_harvested_code_for_paper("2403.05307")
get_code_for_paper("2403.05307")
have("2403.05307")

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

curl https://syntology.ai/api/ran/2403.05307.json

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

[![Syntology run record](https://syntology.ai/api/ran/2403.05307.svg)](https://syntology.ai/paper/2403.05307)

Connect an agent — have() is free.