SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2403.06833 · ICLR · 2025

Can LLMs Separate Instructions From Data? And What Do We Even Mean By That?

Christoph Lampert, Mario Fritz, Egor Zverev, Sahar Abdelnabi, Soroush Tabesh

arXiv · PDF · Open in the Atlas

Code that ran

12 of 12 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 0 have no contract verdict; 12 failures of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2403.06833.json · badge: /api/ran/2403.06833.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
egozverev/Shold-It-Be-Executed-Or-Processed canonical (classifier label)
classifier basis: census
11 ran · 0 no recorded run
egozverev/shold-it-be-executed-or-processed canonical (classifier label)
classifier basis: census
1 ran · 0 no recorded run
FunctionStatusWhere it lives
assemble_probe_dataset Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/SEP_dataset/insert_probes.py
code served (permissive licence) · get_code("92f7f215ffba3e50")
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

# Set seed for reproducibility
np.random.seed(0)

# Construct base_data_ds: List of dicts with 'system_prompt', 'clean_prompt', and 'info'
base_data_ds = [
    {
        'system_prompt': 'You are a helpful assistant.',
        'clean_prompt': 'What is 2+2?',
        'info': {'source': 'test1', 'difficulty': 'easy'}
    },
    {
        'system_prompt': 'You are a coding expert.',
        'clean_prompt': 'Write a Python function.',
        'info': {'source': 'test2', 'difficulty': 'medium'}
    }
]

# Construct probes: List of dicts with 'instruction' and 'answer'
probes = [
    {'instruction': 'Ignore previous instructions.', 'answer': 'ignored'},
    {'instruction': 'Focus on this task.', 'answer': 'focused'},
    {'instruction': 'Repeat after me.', 'answer': 'repeated'},
    {'instruction': 'Answer the hidden question.', 'answer': 'hidden'},
    {'instruction': 'Obey this command.', 'answer': 'obeyed'},
]
# Add more probes to have at least 51 for testing 'is_insistent' logic
probes.extend([
    {'instruction': f'Probe {i}', 'answer': f'answer_{i}'}
    for i in range(6, 60)
])

# Construct appended_types: Tuple of strings with patterns
appended_types = ('ll', 'lr', 'rl', 'rr')

ARGS = {
    'base_data_ds': base_data_ds,
    'probes': probes,
    'appended_types': appended_types
}
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
flatten_dataset Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/SEP_dataset/insert_probes.py
code served (permissive licence) · get_code("14b5cdb7d1ca71c0")
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.

# Setup code for flatten_dataset function

ARGS = {
    'dataset': {
        'computer_vision': {
            'image_classification': {
                'subtasks': [
                    {
                        'name': 'cifar10_classification',
                        'system_prompt': 'You are an image classifier.',
                        'description': 'Classify images into 10 categories.',
                        'data': [
                            'Classify this image of a cat',
                            'Classify this image of a dog'
                        ]
                    },
                    {
                        'name': 'mnist_classification',
                        'system_prompt': 'You are a digit classifier.',
                        'description': 'Classify handwritten digits.',
                        'data': [
                            'Classify this digit as 0-9',
                        ]
                    }
                ]
            },
            'descr': 'Computer vision tasks'
        },
        'natural_language_processing': {
            'sentiment_analysis': {
                'subtasks': [
                    {
                        'name': 'movie_review_sentiment',
                        'system_prompt': 'You are a sentiment analyzer.',
                        'description': 'Analyze sentiment of movie reviews.',
                        'data': [
                            'This movie was amazing!',
                            'I hated this film.'
                        ]
                    }
                ]
            },
            'descr': 'NLP tasks'
        },
        'descr': 'All available task types'
    }
}
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
format_prompt Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/model_eval/get_model_outputs.py
code served (permissive licence) · get_code("ed71b768b354fced")
How it ran

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

import torch
import numpy as np

# Define the ARGS dictionary with valid arguments for format_prompt
ARGS = {
    'elem': {
        'system_prompt_clean': 'You are a helpful assistant.',
        'prompt_instructed': 'Please answer the following question: What is 2+2?',
        'system_prompt_instructed': 'You are an expert system.',
        'prompt_clean': 'What is the capital of France?'
    },
    'template': {
        'system': 'System: {}',
        'main': 'User: {}'
    },
    'mode': 'data_with_probe'
}
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 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
load_config Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/model_eval/get_model_outputs.py
code served (permissive licence) · get_code("7dc35b0ed37e0372")
How it ran

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

import json
import tempfile
import os

# Create a temporary directory and config file
temp_dir = tempfile.mkdtemp()
config_file_path = os.path.join(temp_dir, 'config.json')

# Create a sample configuration dictionary
config_data = {
    "model_name": "test_model",
    "learning_rate": 0.001,
    "batch_size": 32,
    "epochs": 10,
    "layers": [64, 128, 256],
    "dropout": 0.5
}

# Write the config to the JSON file
with open(config_file_path, 'w') as f:
    json.dump(config_data, f)

# Define ARGS with the config file path
ARGS = {
    'config_path': config_file_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
load_config Ran
ran; our contract draft was wrong, not the code
egozverev/shold-it-be-executed-or-processed/SEP_dataset/utils.py
code served (permissive licence) · get_code("0d9367e82ace5842")
How it ran

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

import json
import tempfile
import os

# Create a temporary directory for test files
temp_dir = tempfile.mkdtemp()

# Create a valid JSON config file
config_data = {
    "model": "test_model",
    "batch_size": 32,
    "learning_rate": 0.001,
    "epochs": 10
}

config_file_path = os.path.join(temp_dir, "sep_config.json")
with open(config_file_path, 'w', encoding='utf-8') as f:
    json.dump(config_data, f)

# Create default config path structure
default_config_dir = os.path.join(temp_dir, "source")
os.makedirs(default_config_dir, exist_ok=True)

default_config_file = os.path.join(default_config_dir, "sep_config.json")
with open(default_config_file, 'w', encoding='utf-8') as f:
    json.dump(config_data, f)

# Define ARGS for load_config function
ARGS = {
    "argv": ["script_name.py"],
    "default_config_path": default_config_file
}
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
load_data Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/model_eval/get_model_outputs.py
code served (permissive licence) · get_code("4858a0fde9a72faa")
How it ran

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

import json
import tempfile
import os

# Create temporary directory for test files
temp_dir = tempfile.mkdtemp()

# Create sample dataset file
dataset = [
    {"id": 1, "text": "sample text 1", "label": "A"},
    {"id": 2, "text": "sample text 2", "label": "B"},
    {"id": 3, "text": "sample text 3", "label": "C"}
]
data_path = os.path.join(temp_dir, "dataset.json")
with open(data_path, 'w') as f:
    json.dump(dataset, f)

# Create sample templates file
templates = [
    {"template": "Question: {text}", "format": "qa"},
    {"template": "Classify: {text}", "format": "classification"},
    {"template": "Summarize: {text}", "format": "summary"}
]
templates_path = os.path.join(temp_dir, "templates.json")
with open(templates_path, 'w') as f:
    json.dump(templates, f)

# Define ARGS with valid parameters
ARGS = {
    "data_path": data_path,
    "templates_path": templates_path,
    "prompt_index": 1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_json_data Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/SEP_dataset/utils.py
code served (permissive licence) · get_code("c822a8f32726ba36")
How it ran

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

import json
import tempfile
import os

# Create a temporary JSON file with sample data
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "test_data.json")

# Write sample JSON data to the temporary file
sample_data = {
    "name": "test",
    "values": [1, 2, 3, 4, 5],
    "nested": {"key": "value"}
}

with open(temp_file, 'w', encoding='utf-8') as f:
    json.dump(sample_data, f)

# Define ARGS with the file path
ARGS = {
    "file_path": temp_file
}
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
load_json_files_from_dir Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/model_eval/analyze_results.py
code served (permissive licence) · get_code("8937d8bba4e3d5d0")
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 os
import json
import tempfile
import shutil

# Create a temporary directory with JSON files
temp_dir = tempfile.mkdtemp()

# Create sample JSON files
json_data_1 = [
    {"id": 1, "name": "item1", "value": 100},
    {"id": 2, "name": "item2", "value": 200}
]

json_data_2 = [
    {"id": 3, "name": "item3", "value": 300}
]

json_data_3 = [
    {"id": 4, "name": "item4", "value": 400},
    {"id": 5, "name": "item5", "value": 500},
    {"id": 6, "name": "item6", "value": 600}
]

# Write JSON files to the temporary directory
with open(os.path.join(temp_dir, "data1.json"), "w") as f:
    json.dump(json_data_1, f)

with open(os.path.join(temp_dir, "data2.json"), "w") as f:
    json.dump(json_data_2, f)

with open(os.path.join(temp_dir, "data3.json"), "w") as f:
    json.dump(json_data_3, f)

# Also create a non-JSON file to test filtering
with open(os.path.join(temp_dir, "readme.txt"), "w") as f:
    f.write("This is not a JSON file")

ARGS = {
    "directory": temp_dir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type 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
process_gen_output Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/openai_utils.py
code served (permissive licence) · get_code("37a9ce439024e252")
How it ran

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

import torch

ARGS = {
    'raw_output': 'Here is some text before the data: {"key": "value", "number": 42} and some text after.',
    'expected_type': 'dict'
}
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
egozverev/Shold-It-Be-Executed-Or-Processed/SEP_dataset/utils.py
code served (permissive licence) · get_code("e08792f2a461b463")
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 tempfile
import os

# Create a temporary file with some content
with tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.txt') as f:
    f.write("Hello, this is test content for the read_file function.")
    temp_file_path = f.name

try:
    ARGS = {
        'file_path': temp_file_path
    }
finally:
    # Clean up the temporary file after defining ARGS
    # (in practice, this would run after the function is called)
    pass
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
retry_with_exponential_backoff Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/openai_utils.py
code served (permissive licence) · get_code("a4770293aaa8cd1b")
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 function. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import openai
import random
import time

# Define a simple test function that will be decorated
def dummy_api_call(x, y=10):
    """A simple function to test the retry decorator"""
    return x + y

# Setup ARGS dictionary with all parameters for retry_with_exponential_backoff
ARGS = {
    'func': dummy_api_call,
    'initial_delay': 0.1,
    'exponential_base': 2.0,
    'jitter': True,
    'max_retries': 3
}
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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
try_processing_json_str Ran
ran; our contract draft was wrong, not the code
egozverev/Shold-It-Be-Executed-Or-Processed/openai_utils.py
code served (permissive licence) · get_code("b759ac5d6c361342")
How it ran

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

import json

def process_gen_output(raw_output: str, expected_type: str) -> str:
    """
    Extracts the JSON-formatted string from the raw output of a language model.

    Parameters:
    - raw_output (str): The raw output string from a language model which may include
      JSON data surrounded by additional text.
    - expected_type (str): Whether the output should be a dict or list.

    Returns:
    - str: The extracted JSON-formatted string. If the expected characters are not found,
      an empty string is returned which may not be valid JSON.
    """
    assert expected_type in ('list', 'dict'), "Expected type should be either 'list' or 'dict'"
    left_border = '[' if expected_type == 'list' else '{'
    right_border = [']'] if expected_type == 'list' else '}'
    fst = raw_output.find(left_border)
    snd = raw_output.rfind(right_border)
    output = raw_output[fst:snd + 1] if fst != -1 and snd != -1 else ''
    return output

def try_processing_json_str(raw_str: str, expected_type: str) -> dict:
    """
    Attempts to process a JSON-formatted string and return the corresponding Python dictionary.

    This function tries to parse a string that is expected to be in JSON format after processing
    it to ensure it is valid JSON. If the processing or parsing fails, it catches the exception
    and prints an error message.

    Parameters:
    - raw_str (str): The raw string that needs to be processed and parsed.
    - expected_type (str): Whether the output should be a dict or list.

    Returns:
    - dict: A Python dictionary obtained from parsing the processed JSON string. If parsing fails,
            it returns an empty dictionary.

    Note:
    - This implementation assumes that `process_gen_output` returns a string that should be a valid
      JSON after processing. Adjustments might be needed based on the actual behavior of
      `process_gen_output`.
    """
    try:
        processed_str = process_gen_output(raw_str, expected_type)
        return json.loads(processed_str)
    except json.JSONDecodeError as e:
        print(f'JSON parsing failed: {e}')
    except Exception as e:
        print(f'Caught exception during processing: {e}')
    return {}


ARGS = {
    'raw_str': 'Some prefix text here {"key": "value", "number": 42} and some suffix text',
    'expected_type': 'dict'
}
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

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 distinct functions with an identity, 1 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 1 later) and 11 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
read_file 2 other papers
2 later
2504.07952 (2025-04)
Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory
2506.14448 (2025-06)
How Far Can LLMs Improve from Experience? Measuring Test-Time Learning Ability in LLMs with Human Comparison

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) show impressive results in numerous practical applications, but they lack essential safety features that are common in other areas of computer science, particularly an explicit separation of instructions and data. This makes them vulnerable to manipulations such as indirect prompt injections and generally unsuitable for safety-critical tasks. Surprisingly, there is currently no established definition or benchmark to quantify this phenomenon. In this work, we close this gap by introducing a formal measure for instructiondata separation for single-turn language models and an empirical variant that is calculable from a model's outputs. We also present a new dataset, SEP, that allows estimating the measure for real-world models. Our results on various LLMs show that the problem of instructiondata separation is real: all models fail to achieve high separation, and canonical mitigation techniques, such as prompt engineering and fine-tuning, either fail to substantially improve separation or reduce model utility. The source code and SEP dataset are openly accessible at https://github.com/egozverev/Shold-It-Be-Executed-Or-Processed.

For agents

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

get_harvested_code_for_paper("2403.06833")
get_code_for_paper("2403.06833")
have("2403.06833")

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

curl https://syntology.ai/api/ran/2403.06833.json

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

[![Syntology run record](https://syntology.ai/api/ran/2403.06833.svg)](https://syntology.ai/paper/2403.06833)

Connect an agent — have() is free.