SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2409.03797 · 2024

NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls

arXiv · PDF · Open in the Atlas

Code that ran

17 of 21 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; 7 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 10 failures of our instrument; 4 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2409.03797.json · badge: /api/ran/2409.03797.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
ibm/nestful canonical (classifier label)
classifier basis: census
17 ran · 4 no recorded run
FunctionStatusWhere it lives
ast_parse Ran
ran; our contract draft was wrong, not the code
ibm/nestful/src/ast_utils.py
code served (permissive licence) · get_code("9c6d8f0b3847114f")
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

ARGS = {
    'input_str': "func(arg1=42, arg2='hello', arg3=[1, 2, 3])",
    'language': 'Python'
}
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
create_service_account Ran
ran; our contract draft was wrong, not the code
ibm/nestful/data_v2/executable_functions/py_code_file_1001.py
code served (permissive licence) · get_code("ce18037f4b6348c6")
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)

# Analyzing the function:
# - create_service_account takes company_name: Union[str, bytes]
# - The function calls .replace(' ', '_') on company_name, which works on both str and bytes
# - Then concatenates with '_service'
# - Returns a str

# Test with a string argument
ARGS = {
    "company_name": "Acme Corporation"
}
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
do_stuff Ran
ran, honoured the contract we drafted
ibm/nestful/data_v2/executable_functions/py_code_file_1002.py
code served (permissive licence) · get_code("153e1cc7a8f8d341")
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

torch.manual_seed(0)

ARGS = {
    "x": 5,
    "y": 3.5,
}
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
extend_num Ran
ran, honoured the contract we drafted
ibm/nestful/data_v2/executable_functions/py_code_file_100.py
code served (permissive licence) · get_code("31c2e27c79774797")
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 int, ndim 0, dtype family int, 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

# Set seed for reproducibility
torch.manual_seed(0)

# Construct arguments for extend_num function
# The function takes two numbers (int or float) and returns their sum

# Test case 1: both integers
num1_int = 5
num2_int = 3

# Test case 2: both floats
num1_float = 2.5
num2_float = 1.3

# Test case 3: mixed int and float
num1_mixed = 10
num2_mixed = 4.7

# Using the first test case (both integers)
ARGS = {
    "num1": num1_int,
    "num2": num2_int
}
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 int, ndim 0, dtype family int, 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
find_strings_with_pattern Ran
ran, honoured the contract we drafted
ibm/nestful/data_v2/executable_functions/py_code_file_1.py
code served (permissive licence) · get_code("47cc520d6946b782")
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 int; 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 typing import List

def find_strings_with_pattern(string_list: List[str], pattern: str) -> List[int]:
    """Finds the indices of strings in a list that contain a given substring.

    Args:
        string_list: A list of strings to search for the pattern.
        pattern: The substring pattern to search for in the strings.

    Returns:
        A list of indices of the strings in the input list that contain the pattern as a substring.
    """
    result = []
    for i, string in enumerate(string_list):
        if pattern in string:
            result.append(i)
    return result


ARGS = {
    "string_list": ["hello world", "foo bar", "hello there", "python code"],
    "pattern": "hello"
}
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 int — 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
format_padding Ran
ran; our contract draft was wrong, not the code
ibm/nestful/data_v2/executable_functions/py_code_file_10.py
code served (permissive licence) · get_code("fb5af146b3521b50")
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)

# Construct valid arguments for format_padding
string = "hello"
padding_length = 10
padding_char = "*"

ARGS = {
    "string": string,
    "padding_length": padding_length,
    "padding_char": padding_char
}
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
get_deli_sep_str_list Ran
ran; our contract draft was wrong, not the code
ibm/nestful/src/output_parsers.py
code served (permissive licence) · get_code("a58b18082fe1da3f")
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 torch
import numpy as np

ARGS = {
    'text': 'hello, "world, test", foo, bar',
    'deli': ','
}
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 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
get_icl_str Ran
ran; our contract draft was wrong, not the code
ibm/nestful/src/instruct_data_prep.py
code served (permissive licence) · get_code("48207fba685d16d6")
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 json

# Construct icl_examples - a list of dicts with 'input' and 'output' keys
# Each output is a list of function call objects with 'name' field
icl_examples = [
    {
        'input': 'What is the weather today?',
        'output': [
            {'name': 'get_weather', 'params': {'location': 'New York'}},
            {'name': 'get_temperature', 'params': {'unit': 'celsius'}}
        ]
    },
    {
        'input': 'Tell me a joke',
        'output': [
            {'name': 'generate_joke', 'params': {'style': 'pun'}}
        ]
    }
]

# Test with different model names to cover branches
model_name = 'xLAM-7b-fc-r'

ARGS = {
    'icl_examples': icl_examples,
    'model_name': model_name
}
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
listit Ran
ran; our contract draft was wrong, not the code
ibm/nestful/src/scorer.py
code served (permissive licence) · get_code("f0a4f5a7f292a3d0")
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. 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 torch.nn as nn
from typing import Union, List, Tuple

torch.manual_seed(0)

def listit(t):
    return list(map(listit, t)) if isinstance(t, (list, tuple)) else t

# The function recursively converts nested lists/tuples to lists
# It takes a single parameter `t` which can be:
# - A list or tuple (which gets recursively processed)
# - Any other type (which gets returned as-is)

# Build test arguments with nested structures
nested_list = [1, [2, 3], [[4, 5], 6]]
nested_tuple = (1, (2, 3), ((4, 5), 6))
nested_mixed = [1, (2, [3, 4]), 5]
tensor_nested = [torch.tensor([1, 2]), [torch.tensor([3, 4])]]
simple_value = 42

ARGS = {
    "t": nested_list
}
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 list — a list 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
permutation Ran
ran, honoured the contract we drafted
ibm/nestful/data_v2/executable_functions/basic_functions.py
code served (permissive licence) · get_code("b524a104e61bdc60")
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 int, ndim 0, dtype family int, 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
from math import factorial

# Construct arguments for the permutation function
# permutation(n: int, r: int) -> int
# The function calculates permutations using factorial

ARGS = {
    "n": 5,
    "r": 3
}
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 int, ndim 0, dtype family int, 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
power Ran
ran, honoured the contract we drafted
ibm/nestful/data_v2/executable_functions/basic_functions.py
code served (permissive licence) · get_code("29da7d03953ecf87")
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 int, ndim 0, dtype family int, 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 typing

torch_seed = 0

ARGS = {
    "base": 2,
    "exponent": 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 int, ndim 0, dtype family int, 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
process_slot_value Ran
ran; our contract draft was wrong, not the code
ibm/nestful/src/output_parsers.py
code served (permissive licence) · get_code("cd5994aa5f8341d1")
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 process_slot_value(slot_value):
    if slot_value.startswith("'"):
        slot_value = slot_value[1:]
    if slot_value.endswith("'"):
        slot_value = slot_value[:-1]
    if not slot_value.startswith('"'):
        slot_value = '"' + slot_value
    if not slot_value.endswith('"'):
        slot_value = slot_value + '"'
    sub_slot_value = slot_value[1:-1].replace('"', "'")
    return sub_slot_value

ARGS = {
    "slot_value": "'test_value'"
}
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
read_file Ran
ran; our contract draft was wrong, not the code
ibm/nestful/src/utils.py
code served (permissive licence) · get_code("ba6cfefa8094dcb6")
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 tempfile
import os

# Create a temporary file with some test content
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt')
temp_file.write("line 1\n")
temp_file.write("line 2\n")
temp_file.write("line 3\n")
temp_file.close()

ARGS = {
    'path': temp_file.name
}
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
resolve_ast_by_type Ran
ran, honoured the contract we drafted
ibm/nestful/src/ast_utils.py
code served (permissive licence) · get_code("cc91f822de63292f")
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 int, ndim 0, dtype family int, 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 ast
import torch
import numpy as np

# The function resolve_ast_by_type takes a single parameter: value
# which should be an AST node representing some Python expression.
# We need to construct various AST nodes to test the function.

torch.manual_seed(0)

# Create various AST nodes to pass as the 'value' parameter
ast_constant = ast.Constant(value=42)
ast_string = ast.Constant(value="hello")
ast_ellipsis_const = ast.Constant(value=...)
ast_none = ast.Constant(value=None)
ast_true = ast.Constant(value=True)

# UnaryOp (negative number)
ast_unary = ast.UnaryOp(
    op=ast.USub(),
    operand=ast.Constant(value=5)
)

# List
ast_list = ast.List(
    elts=[
        ast.Constant(value=1),
        ast.Constant(value=2),
        ast.Constant(value=3)
    ],
    ctx=ast.Load()
)

# Dict
ast_dict = ast.Dict(
    keys=[ast.Constant(value="key1"), ast.Constant(value="key2")],
    values=[ast.Constant(value=10), ast.Constant(value=20)]
)

# Name (identifier)
ast_name = ast.Name(id="some_variable", ctx=ast.Load())

# BinOp (binary operation)
ast_binop = ast.BinOp(
    left=ast.Constant(value=10),
    op=ast.Add(),
    right=ast.Constant(value=5)
)

# Tuple
ast_tuple = ast.Tuple(
    elts=[
        ast.Constant(value=1),
        ast.Constant(value=2)
    ],
    ctx=ast.Load()
)

# Call without keywords
ast_call_no_kwargs = ast.Call(
    func=ast.Name(id="func", ctx=ast.Load()),
    args=[ast.Constant(value=1)],
    keywords=[]
)

# Call with keywords (for resolve_ast_call)
ast_call_with_kwargs = ast.Call(
    func=ast.Attribute(
        value=ast.Name(id="obj", ctx=ast.Load()),
        attr="method",
        ctx=ast.Load()
    ),
    args=[],
    keywords=[
        ast.keyword(arg="param1", value=ast.Constant(value=42)),
        ast.keyword(arg="param2", value=ast.Constant(value="test"))
    ]
)

# Ellipsis
ast_ellipsis = ast.Ellipsis()

# Subscript
ast_subscript = ast.Subscript(
    value=ast.Name(id="arr", ctx=ast.Load()),
    slice=ast.Constant(value=0),
    ctx=ast.Load()
)

ARGS = {
    'value': ast_constant
}
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 int, ndim 0, dtype family int, 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
resolve_ast_call Ran
ran; our contract draft was wrong, not the code
ibm/nestful/src/ast_utils.py
code served (permissive licence) · get_code("053e58f476c2e09c")
How it ran

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

import ast

# Create a simple AST element representing a function call with keyword arguments
# The function looks at: elem.func, elem.keywords
# For elem.func, it traverses Attribute nodes and finds the final Name node
# For elem.keywords, it processes arg.arg (name) and arg.value (the AST node)

# Build an AST Call node: example_module.example_func(param1=42, param2="test")
call_node = ast.Call(
    func=ast.Attribute(
        value=ast.Name(id='example_module', ctx=ast.Load()),
        attr='example_func',
        ctx=ast.Load()
    ),
    args=[],
    keywords=[
        ast.keyword(arg='param1', value=ast.Constant(value=42)),
        ast.keyword(arg='param2', value=ast.Constant(value="test"))
    ]
)

ARGS = {
    'elem': call_node
}
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
simplify_slug Ran
ran; our contract draft was wrong, not the code
ibm/nestful/data_v2/executable_functions/py_code_file_1000.py
code served (permissive licence) · get_code("662385c90863ea91")
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
from typing import Union

def simplify_slug(slug: Union[str, None]) -> str:
    """Converts a string into a simplified slug.

    The slug is lowercase and contains only alphanumeric characters and hyphens.

    Removes any leading or trailing whitespace from the input string and replaces

    any sequences of whitespace with a single hyphen. If the resulting slug is

    empty, returns 'slug-not-provided'.

    Args:

        slug: The input string to be converted into a slug.

    Returns:

        The simplified slug.
    """
    slug = re.sub('\\s+', '-', slug.lower().strip())
    slug = re.sub('[^a-z0-9-]', '', slug)
    if not slug:
        return 'slug-not-provided'
    return slug


# Test cases for different scenarios
ARGS = {
    "slug": "Hello World Test"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
subtract Ran
ran, honoured the contract we drafted
ibm/nestful/data_v2/executable_functions/basic_functions.py
code served (permissive licence) · get_code("7f2f2e2df15e6716")
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 int, ndim 0, dtype family int, 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

torch.manual_seed(0)

ARGS = {
    "arg_0": 10,
    "arg_1": 3
}
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 int, ndim 0, dtype family int, 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
calculate_ans Not yet run
unverified (harvested, no recorded run)
ibm/nestful/src/scorer.py
code served (permissive licence) · get_code("f5b3c135f40d531c")
How it ran

no invocation recorded.

get_basic_func_list Not yet run
unverified (harvested, no recorded run)
ibm/nestful/src/scorer.py
code served (permissive licence) · get_code("c9fb850a9577a4d4")
How it ran

no invocation recorded.

get_granite_tokenizer Not yet run
unverified (harvested, no recorded run)
ibm/nestful/src/instruct_data_prep.py
code served (permissive licence) · get_code("63f870a98c11420f")
How it ran

no invocation recorded.

granite_3_1_prompt_input Not yet run
unverified (harvested, no recorded run)
ibm/nestful/src/instruct_data_prep.py
code served (permissive licence) · get_code("0d5485e114c8ba34")
How it ran

no invocation recorded.

Also present in other papers' code

Of 21 samples lifted from this paper's linked repositories, 21 carry a function identity and 0 do not. Of the 21 distinct functions with an identity, 21 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 21 later) and 0 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
resolve_ast_by_type 2 other papers
2 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
2507.01299 (2025-07)
La RoSA: Enhancing LLM Efficiency via Layerwise Rotated Sparse Activation
ast_parse 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
calculate_ans 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
create_service_account 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
do_stuff 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
extend_num 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
find_strings_with_pattern 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
format_padding 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
get_basic_func_list 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
get_deli_sep_str_list 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
get_granite_tokenizer 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
get_icl_str 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
granite_3_1_prompt_input 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
listit 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
permutation 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
power 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
process_slot_value 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
read_file 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
resolve_ast_call 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls
simplify_slug 1 other paper
1 later
2025.emnlp-main.1702 (2025 (year only))
NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls

Shown: the 20 functions present in the most other papers. 1 more shared functions are not listed here.

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

The resurgence of autonomous agents built using large language models (LLMs) to solve complex real-world tasks has brought increased focus on LLMs' fundamental ability of tool or function calling. At the core of these agents, an LLM must plan, execute, and respond using external tools, APIs, and custom functions. Research on tool calling has gathered momentum, but evaluation benchmarks and datasets representing the complexity of the tasks have lagged behind. In this work, we focus on one such complexity, nested sequencing, with the goal of extending existing benchmarks and evaluation. Specifically, we present NESTFUL, a benchmark to evaluate LLMs on nested sequences of API calls, i.e., sequences where the output of one API call is passed as input to a subsequent call. NESTFUL contains 1800+ nested sequences where all the function calls are executable. Experimental results on a variety of models show that the best-performing model (GPT-4o) achieves a full sequence match accuracy of 28% and a win-rate of 60%, necessitating a large scope for improvement in the nested sequencing aspect of function calling. Our analysis of these results provides possible future research directions for the community, in addition to a benchmark to track progress. We have released the NESTFUL dataset under the Apache 2.0 license at https://github.com/IBM/NESTFUL.

For agents

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

get_harvested_code_for_paper("2409.03797")
get_code_for_paper("2409.03797")
have("2409.03797")

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

curl https://syntology.ai/api/ran/2409.03797.json

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

[![Syntology run record](https://syntology.ai/api/ran/2409.03797.svg)](https://syntology.ai/paper/2409.03797)

Connect an agent — have() is free.