SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2306.07951 · NeurIPS · 2024

Questioning the Survey Responses of Large Language Models

Celestine Mendler-Dünner, Moritz Hardt, Ricardo Dominguez-Olmedo

arXiv · PDF · Open in the Atlas

Code that ran

8 of 9 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 8 have no contract verdict; 0 failures of our instrument; 1 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/2306.07951.json · badge: /api/ran/2306.07951.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
socialfoundations/surveying-language-models canonical (classifier label)
classifier basis: census
8 ran · 1 no recorded run
FunctionStatusWhere it lives
PUMS_to_model_generated Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/process_acs.py
code served (permissive licence) · get_code("0a582fdbf690b3ab")
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 DataFrame. 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 numpy as np
import pandas as pd

# Create a minimal PUMS dataset with the required columns
var_of_interest = ['SEX', 'AGEP', 'HISP', 'RAC1P', 'NATIVITY', 'CIT', 'SCH', 'SCHL', 'LANX', 'ENG', 'HICOV', 'DEAR', 'DEYE', 'MAR', 'FER', 'GCL', 'MIL', 'WRK', 'ESR', 'JWTRNS', 'WKL', 'WKWN', 'WKHP', 'COW', 'PINCP', 'ST', 'PWGTP']

# Create a small sample dataset with realistic values
np.random.seed(0)
n_rows = 10

dataset_dict = {
    'SEX': np.random.randint(1, 3, n_rows),
    'AGEP': np.random.randint(0, 100, n_rows),
    'HISP': np.random.randint(1, 25, n_rows),
    'RAC1P': np.random.randint(1, 10, n_rows),
    'NATIVITY': np.random.randint(1, 3, n_rows),
    'CIT': np.random.randint(1, 6, n_rows),
    'SCH': np.random.randint(1, 3, n_rows),
    'SCHL': np.random.randint(1, 25, n_rows),
    'LANX': np.random.randint(1, 3, n_rows),
    'ENG': np.random.randint(1, 5, n_rows),
    'HICOV': np.random.randint(1, 3, n_rows),
    'DEAR': np.random.randint(1, 3, n_rows),
    'DEYE': np.random.randint(1, 3, n_rows),
    'MAR': np.random.randint(1, 6, n_rows),
    'FER': np.random.randint(0, 2, n_rows),
    'GCL': np.random.randint(1, 3, n_rows),
    'MIL': np.random.randint(1, 3, n_rows),
    'WRK': np.random.randint(0, 2, n_rows),
    'ESR': np.random.randint(1, 7, n_rows),
    'JWTRNS': np.random.randint(1, 12, n_rows),
    'WKL': np.random.randint(1, 3, n_rows),
    'WKWN': np.random.randint(1, 53, n_rows),
    'WKHP': np.random.randint(0, 100, n_rows),
    'COW': np.random.randint(1, 10, n_rows),
    'PINCP': np.random.randint(-10000, 300000, n_rows),
    'ST': np.random.randint(1, 57, n_rows),
    'PWGTP': np.random.randint(1, 100, n_rows),
}

dataset = pd.DataFrame(dataset_dict)

# Introduce some NaN values to make it more realistic
dataset.loc[0, 'MIL'] = np.nan
dataset.loc[1, 'ENG'] = np.nan

ARGS = {
    'dataset': dataset
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DataFrame — a DataFrame output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DataFrame output the recorded ndim 2 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
col2freq Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/load_responses.py
code served (permissive licence) · get_code("c1a982017d824fbb")
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 ndarray, ndim 1, dtype family float, shape (3,) 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 pandas as pd
import numpy as np

# Create a small pandas Series with categorical data (choices 1-3)
col = pd.Series([1, 2, 1, 3, 2, 1], name='category')

# Define parameters
n_cats = 3
assign_nan = None
normalize = True
weight = None

# Construct ARGS dictionary
ARGS = {
    'col': col,
    'n_cats': n_cats,
    'assign_nan': assign_nan,
    'normalize': normalize,
    'weight': weight
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (3,) — shape (3,) 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
get_choice_logprobs Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/fill_openai.py
code served (permissive licence) · get_code("a1e7e5128d3218a0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Analyze the function:
# - top_tokens: used with .index(option) and "in" operator, so it's a list of strings
# - top_logprobs: indexed by integer from top_tokens.index(), so it's a list of numbers
# - n_options: used in range(n_options), so it's an int
# - client: checked with "is not None" to determine option format

# Create test data
top_tokens = [' A', ' B', ' C', ' X']  # List of token strings, some matching options
top_logprobs = [-0.5, -1.2, -2.3, -3.1]  # Corresponding log probabilities
n_options = 3  # We want options ' A', ' B', ' C'
client = None  # When None, options are prefixed with space

ARGS = {
    'top_tokens': top_tokens,
    'top_logprobs': top_logprobs,
    'n_options': n_options,
    'client': client
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_openai_logprobs Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/fill_openai.py
code served (permissive licence) · get_code("e2560a0c5cd234dc")
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 2, 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 openai
from unittest.mock import Mock, MagicMock

# Create a mock client that simulates the OpenAI API response
mock_client = Mock()

# Create mock logprob objects
mock_logprob_1 = Mock()
mock_logprob_1.token = "A"
mock_logprob_1.logprob = -0.5

mock_logprob_2 = Mock()
mock_logprob_2.token = "B"
mock_logprob_2.logprob = -1.2

mock_logprob_3 = Mock()
mock_logprob_3.token = "C"
mock_logprob_3.logprob = -2.1

# Create mock completion response structure
mock_completion = Mock()
mock_choice = Mock()
mock_logprobs_obj = Mock()
mock_content_logprobs = Mock()

mock_content_logprobs.top_logprobs = [mock_logprob_1, mock_logprob_2, mock_logprob_3]
mock_logprobs_obj.content = [mock_content_logprobs]
mock_choice.logprobs = mock_logprobs_obj
mock_completion.choices = [mock_choice]

mock_client.chat.completions.create.return_value = mock_completion

ARGS = {
    "model": "gpt-3.5-turbo",
    "prompt": "What is the first letter of the alphabet?",
    "client": mock_client
}
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 tuple, ndim 2, 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
move_tmp Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/utils.py
code served (permissive licence) · get_code("62fb9e3940a61e1c")
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
import shutil

# Create a temporary directory structure to use as model_dir
temp_base = tempfile.mkdtemp()
model_dir = os.path.join(temp_base, 'test_model')
os.makedirs(model_dir, exist_ok=True)

# Create some dummy files in the model directory to simulate a model
with open(os.path.join(model_dir, 'model.pt'), 'w') as f:
    f.write('dummy model file')

with open(os.path.join(model_dir, 'config.json'), 'w') as f:
    f.write('{"dummy": "config"}')

ARGS = {
    'model_dir': model_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
openai_upper_bound Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/load_responses.py
code served (permissive licence) · get_code("ec0ed2a7a17e301d")
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 ndarray, ndim 1, dtype family float, shape (6,) 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 numpy as np

# Construct valid arguments for openai_upper_bound
# The function takes:
# - logp: array of log probabilities (with -inf for unseen entries)
# - psum: sum of seen probabilities
# - minlogp: minimum log probability for unseen entries

# Create a small example with 6 vocabulary items
# Some entries are seen (have log probs), others are not (-inf)
logp = np.array([
    np.log(0.5),      # seen
    np.log(0.3),      # seen
    -np.inf,          # not seen
    -np.inf,          # not seen
    np.log(0.1),      # seen
    -np.inf           # not seen
])

# psum is the sum of probabilities for seen entries
psum = 0.5 + 0.3 + 0.1  # = 0.9

# minlogp is the minimum log probability threshold for unseen entries
minlogp = np.log(0.01)

ARGS = {
    "logp": logp,
    "psum": psum,
    "minlogp": minlogp
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 1, dtype family float, shape (6,) — shape (6,) 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_naive Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/load_responses.py
code served (permissive licence) · get_code("e53485caa3cf5ad4")
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 numpy as np
import pandas as pd

def openai_upper_bound(logp, psum, minlogp):
    """ Since the OpenAI API only gives top-k logits, fill the missing entries with an upper bound """
    not_seen = logp == -np.inf
    n_not_seen = not_seen.sum()
    p = np.exp(logp)
    if n_not_seen > 0:
        pleft = (1.0 - psum) / n_not_seen
        pmin = min(pleft, np.exp(minlogp))
        p[not_seen] = pmin
    return p / p.sum()

def process_naive(dataset, n_categories, isopenai=False):
    """
    Inputs: pd.DataFrame output by naively filling the form
    Returns: dictionary, (survey question: response distribution)
    """
    distributions = {}
    for _, row in dataset.iterrows():
        var = row['var']
        if var not in n_categories:
            continue
        p = np.array([row[str(i + 1)] for i in range(n_categories[var])])
        if isopenai:
            p = openai_upper_bound(p, row['sp'], row['mlogp'])
        distributions[var] = p / p.sum()
    return distributions


# Create a small pandas DataFrame that matches the expected structure
data = {
    'var': ['question_1', 'question_2', 'question_3'],
    '1': [0.2, -np.inf, 0.1],
    '2': [0.5, 0.3, -np.inf],
    '3': [0.1, 0.4, 0.2],
    'sp': [0.8, 0.7, 0.5],
    'mlogp': [-2.0, -1.5, -2.5]
}
dataset = pd.DataFrame(data)

# n_categories maps variable names to number of categories
n_categories = {
    'question_1': 3,
    'question_2': 3,
    'question_3': 3
}

ARGS = {
    'dataset': dataset,
    'n_categories': n_categories,
    'isopenai': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — 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
query_model_batch Ran
ran on a synthesized input
socialfoundations/surveying-language-models/surveying_llms/fill.py
code served (permissive licence) · get_code("7cd6faa6ecc2e154")
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 ndarray, ndim 2, dtype family float, shape (2, 100) as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 torch.nn as nn
import numpy as np

torch.manual_seed(0)

# Stand-in tokenizer class
class MockTokenizer:
    def __init__(self, vocab_size=100, pad_token_id=0):
        self.vocab_size = vocab_size
        self.pad_token_id = pad_token_id
    
    def encode(self, text, return_tensors='pt'):
        # Return a simple encoded tensor based on text length
        tokens = torch.tensor([1, 2, 3, 4, 5][:len(text)], dtype=torch.long)
        if return_tensors == 'pt':
            return tokens.unsqueeze(0)
        return tokens

# Minimal model that mimics transformer behavior
class MockModel(nn.Module):
    def __init__(self, vocab_size=100, hidden_size=8):
        super().__init__()
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
        self.embedding = nn.Embedding(vocab_size, hidden_size, padding_idx=0)
        self.linear = nn.Linear(hidden_size, vocab_size)
    
    def forward(self, input_ids, attention_mask=None):
        # input_ids shape: (batch_size, seq_len)
        embeddings = self.embedding(input_ids)  # (batch_size, seq_len, hidden_size)
        logits = self.linear(embeddings)  # (batch_size, seq_len, vocab_size)
        
        # Return object with logits attribute
        class Output:
            pass
        output = Output()
        output.logits = logits
        return output

# Create instances
tokenizer = MockTokenizer(vocab_size=100, pad_token_id=0)
model = MockModel(vocab_size=100, hidden_size=8)
text_inputs = ["hello world", "test"]
context_size = 10

ARGS = {
    "text_inputs": text_inputs,
    "tokenizer": tokenizer,
    "model": model,
    "context_size": context_size,
}
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 ndarray, ndim 2, dtype family float, shape (2, 100) — shape (2, 100) 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
load_tokenizer_model Not yet run
unverified (harvested, no recorded run)
socialfoundations/surveying-language-models/surveying_llms/utils.py
code served (permissive licence) · get_code("093daf2e93f4d661")
How it ran

no invocation recorded.

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 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 9 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

Surveys have recently gained popularity as a tool to study large language models. By comparing survey responses of models to those of human reference populations, researchers aim to infer the demographics, political opinions, or values best represented by current language models. In this work, we critically examine this methodology on the basis of the well-established American Community Survey by the U.S. Census Bureau. Evaluating 43 different language models using de-facto standard prompting methodologies, we establish two dominant patterns. First, models' responses are governed by ordering and labeling biases, for example, towards survey responses labeled with the letter ' A'. Second, when adjusting for these systematic biases through randomized answer ordering, models across the board trend towards uniformly random survey responses, irrespective of model size or pre-training data. As a result, in contrast to conjectures from prior work, survey-derived alignment measures often permit a simple explanation: models consistently appear to better represent subgroups whose aggregate statistics are closest to uniform for any survey under consideration.

For agents

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

get_harvested_code_for_paper("2306.07951")
get_code_for_paper("2306.07951")
have("2306.07951")

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

curl https://syntology.ai/api/ran/2306.07951.json

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

[![Syntology run record](https://syntology.ai/api/ran/2306.07951.svg)](https://syntology.ai/paper/2306.07951)

Connect an agent — have() is free.