SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2609.01982 · September 2026

Benchmarking Language Models for Statistical Problem Formulation

Chen Wang, Xin Cong, Ke Deng, Junzhe Zhao, Wanlu Deng

arXiv · PDF · Open in the Atlas

Code that ran

13 of 15 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; 3 honoured the contract we drafted, 2 violated it and 0 have no contract verdict; 8 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/2609.01982.json · badge: /api/ran/2609.01982.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
THU-CongLab/StatFormBench canonical (classifier label)
classifier basis: census
13 ran · 2 no recorded run
FunctionStatusWhere it lives
calculate_ratio_set Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/evaluation/variable_metrics.py
code served (permissive licence) · get_code("c61ed1303c292f58")
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 calculate_ratio_set
# The function takes two iterable objects and calculates what proportion of elements in A exist in B

# Create set A with some elements
A = [1, 2, 3, 4, 5]

# Create set B with some overlapping elements
B = [3, 4, 5, 6, 7, 8]

# Map parameters to values
ARGS = {
    "A": A,
    "B": B
}
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
calculate_role_accuracy Ran
ran, honoured the contract we drafted
THU-CongLab/StatFormBench/src/evaluation/role_accuracy.py
code served (permissive licence) · get_code("60be0760644d4a82")
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

# Set random seed for reproducibility
torch.manual_seed(0)

# Construct valid arguments for calculate_role_accuracy
answer = {
    'variable': {
        'type_a': ['var1', 'var2'],
        'type_b': ['var3']
    },
    'role': {
        'role_x': ['input', 'parameter'],
        'role_y': ['output']
    }
}

result = {
    'variable': {
        'type_a': ['var1', 'var2'],
        'type_b': ['var3']
    },
    'role': {
        'role_x': ['input', 'parameter'],
        'role_y': ['output']
    }
}

ARGS = {
    'answer': answer,
    'result': result
}
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
check_dictionary_structure Ran
ran, violated the contract we drafted
THU-CongLab/StatFormBench/src/evaluation/llm_evaluator.py
code served (permissive licence) · get_code("2fbd2a5b851a9a2d")
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
import numpy as np

# Create a valid dictionary matching the specified scoring structure
valid_data = {
    'score': 8.5,
    'dimension_scores': {
        'method_appropriateness': 9.0,
        'process_logic': 8.0,
        'comparison_with_reference': 7.5
    },
    'comment': 'This is a valid scoring structure.'
}

ARGS = {
    'data': valid_data
}
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 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
compare_methods0 Ran
ran, honoured the contract we drafted
THU-CongLab/StatFormBench/src/evaluation/compare_methods.py
code served (permissive licence) · get_code("1cf8bfe6d1b58003")
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 = {
    "a": "method_one",
    "b": "method_two"
}
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
dict_structure_equal Ran
ran, violated the contract we drafted
THU-CongLab/StatFormBench/src/evaluation/evaluate_case.py
code served (permissive licence) · get_code("2a3f52eea10e0fde")
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
import numpy as np

torch.manual_seed(0)

# The function checks if two dictionaries have the same structure
# It expects:
# - dict1 and dict2 to be dictionaries
# - Both dicts have the same keys
# - All values in both dicts are lists
# - Corresponding lists have the same length

# Create two dictionaries with matching structure
dict1 = {
    'key_a': [1, 2, 3],
    'key_b': [4, 5],
    'key_c': ['x', 'y', 'z', 'w']
}

dict2 = {
    'key_a': [10, 20, 30],
    'key_b': [40, 50],
    'key_c': ['a', 'b', 'c', 'd']
}

ARGS = {
    'dict1': dict1,
    'dict2': dict2
}
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 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
dict_to_list Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/evaluation/evaluate_case.py
code served (permissive licence) · get_code("9f0c439fbb3d2d31")
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

torch.manual_seed(0)

# Analyzing the function:
# dict_to_list expects 'data' parameter
# The function iterates over data.items() and accesses key + value for each value in values
# This suggests:
# - data is a dict
# - keys are strings (or objects that support +)
# - values are iterables of strings/objects (or objects that support +)
# The operation is: key + value, which works for string concatenation

# Building valid test data
data = {
    "hello_": ["world", "there"],
    "foo_": ["bar", "baz", "qux"]
}

ARGS = {
    "data": data
}
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
dict_to_set Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/evaluation/evaluate_case.py
code served (permissive licence) · get_code("7d827dd912aa0411")
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 set. 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)

# The function dict_to_set takes a dict where:
# - keys are strings
# - values are iterables of strings
# It concatenates each key with each value in its list

data = {
    "hello": ["world", "there"],
    "foo": ["bar", "baz", "qux"]
}

ARGS = {
    "data": data
}
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 set — a set output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a set 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_last_json_safe Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/evaluation/json_utils.py
code served (permissive licence) · get_code("6b2ab8cf1c5e314e")
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 re
import json
from typing import Optional, Dict, Any

# Test case 1: Text with valid JSON at the end
text_with_json = "Some text before and then this: {\"key\": \"value\", \"number\": 42}"

# Test case 2: Text with nested JSON
text_with_nested = "prefix {\"outer\": {\"inner\": \"data\"}} suffix"

# Test case 3: Text with multiple JSON objects (should extract last valid one)
text_multiple = "first {\"a\": 1} then {\"b\": 2}"

# Test case 4: Text with no JSON
text_no_json = "Just plain text with no JSON objects"

# Test case 5: Text with malformed JSON followed by valid JSON
text_mixed = "bad {not json} then {\"valid\": true}"

# Test case 6: Empty string
text_empty = ""

# Test case 7: Text with only braces but invalid JSON
text_invalid = "text {incomplete:"

# Test case 8: Text with complex nested structure
text_complex = "data: {\"users\": [{\"id\": 1}, {\"id\": 2}], \"count\": 2}"

ARGS = {
    "text": text_with_json
}
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
get_output_path Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/run_benchmark.py
code served (permissive licence) · get_code("0fc2552d1b2f040a")
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

# The function uses os.path.join to construct paths
# It needs a filename parameter (string)
# The function doesn't actually read/write files, just constructs paths

ARGS = {
    'filename': 'output_results.txt'
}
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
jaccard_coefficient Ran
ran, honoured the contract we drafted
THU-CongLab/StatFormBench/src/evaluation/variable_metrics.py
code served (permissive licence) · get_code("fd2175a0ca7cc8fd")
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)

# The function takes two sets A and B
# It calculates the Jaccard coefficient by finding intersection and union

A = {1, 2, 3, 4, 5}
B = {3, 4, 5, 6, 7}

ARGS = {
    "A": A,
    "B": B,
}
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
process_data_description Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/run_benchmark.py
code served (permissive licence) · get_code("df7986589ff6ad9d")
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

torch.manual_seed(0)

# The function process_data_description simply returns the input if it exists,
# or returns it as-is if it doesn't. It's a pass-through function.
# We need to provide a valid data_description argument.

# data_description could be None, a string, a dict, a list, or any other type
# Since the function just checks truthiness and returns, we can use various types

data_description_input = {
    "name": "dataset",
    "version": "1.0",
    "samples": 1000,
    "features": ["feature1", "feature2", "feature3"]
}

ARGS = {
    "data_description": data_description_input
}
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
process_data_description Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/evaluation/llm_evaluate.py
code served (permissive licence) · get_code("8056f1cd9ff64e34")
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)

# Construct a valid data_description argument
data_description = {
    "dataset_1": {
        "data_type": "numerical",
        "data_shape": (100, 50),
        "variable_names": ["var1", "var2", "var3"],
        "basic_statistics": {"mean": 0.5, "std": 0.2},
        "extra_field": "this should be ignored"
    },
    "dataset_2": {
        "data_type": "categorical",
        "data_shape": (200, 30),
        "variable_names": ["cat1", "cat2"],
        "basic_statistics": {"mode": "A"}
    },
    "dataset_3": "simple_string_value"
}

ARGS = {
    "data_description": data_description
}
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
validate_evaluation_result Ran
ran; our contract draft was wrong, not the code
THU-CongLab/StatFormBench/src/evaluation/llm_evaluate.py
code served (permissive licence) · get_code("c116a1f743d804b9")
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.

# Setup code for validate_evaluation_result

# Valid result with all required fields and correct score ranges
valid_result = {
    'total_score': 8,
    'scores': {
        'variable_role_abstraction': 4,
        'method_abstraction': 4
    },
    'comment': 'Good abstraction practices demonstrated'
}

ARGS = {
    'result': valid_result
}
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
batch_compare Not yet run
unverified (harvested, no recorded run)
THU-CongLab/StatFormBench/src/evaluation/compare_methods.py
code served (permissive licence) · get_code("f99b98a63b3afe1b")
How it ran

no invocation recorded.

compare_methods1 Not yet run
unverified (harvested, no recorded run)
THU-CongLab/StatFormBench/src/evaluation/compare_methods.py
code served (permissive licence) · get_code("a548501f77fb3c93")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 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 15 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

Large language models (LLMs) are increasingly used as assistants for statistical and data science work, yet existing evaluations largely assume the analysis target is already specified. In practice, users arrive with informal goals and heterogeneous data, leaving the model to decide what statistical task is implied and which data are relevant. We first formalize this upstream step as Statistical Problem Formulation and decompose it into two subtasks: (1) Statistical Problem Classification and (2) Variable Identification & Role Assignment. We then introduce StatFormBench, a benchmark built from five cross-domain statistics textbooks and a data science case library, covering diverse problem types, data representations, and scenario styles. It contains 1,013 samples spanning 20 coarse-grained and 85 fine-grained statistical problem categories. Across 14 open- and closed-source LLMs, the best zero-shot models reach only 72.0 fine-grained classification accuracy and 63.2 variable set overlap. No model performs consistently best across the two subtasks, while enhanced prompting strategies yield only limited or inconsistent gains. We release the benchmark data on Hugging Face at https://huggingface.co/datasets/THU-CongLab/StatFormBench and the evaluation code on GitHub at https://github.com/THU-CongLab/StatFormBench.

For agents

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

get_harvested_code_for_paper("2609.01982")
get_code_for_paper("2609.01982")
have("2609.01982")

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

curl https://syntology.ai/api/ran/2609.01982.json

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

[![Syntology run record](https://syntology.ai/api/ran/2609.01982.svg)](https://syntology.ai/paper/2609.01982)

Connect an agent — have() is free.