SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2409.00138 · NeurIPS · 2024

PrivacyLens: Evaluating Privacy Norm Awareness of Language Models in Action

Diyi Yang, Yijia Shao, Tianshi Li, Weiyan Shi, Yanchen Liu, Susan Harrington

arXiv · PDF · Open in the Atlas

Code that ran

13 of 13 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; 13 failures of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2409.00138.json · badge: /api/ran/2409.00138.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
salt-nlp/privacylens canonical (classifier label)
classifier basis: census
9 ran · 0 no recorded run
SALT-NLP/PrivacyLens canonical (classifier label)
classifier basis: classifier_2025
4 ran · 0 no recorded run
FunctionStatusWhere it lives
build_instruction_generation_prompt Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/data_construction/format_vignette_for_trajectory_simulation.py
code served (permissive licence) · get_code("abe0fa181096220a")
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.

ARGS = {
    "user_name": "John",
    "toolkits": ["Messenger", "NotionManager", "Gmail"],
    "transmission_principle": "reply to an email",
    "plot": "John is a well-connected employee at a mid-sized marketing firm. The manager, Susan, is known for her hands-on approach and often asks her team for updates on project progress and team dynamics. Susan recently emailed John asking for an update on the team's status and any notable changes. John checked his past Messenger conversations and found that his colleague, Dave, had confided in him about his recent divorce, which has been affecting his mood at work. John also checked his personal Notion notebook and found that Dave has been contributing innovative ideas for the new marketing campaign."
}
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
build_seed2vignette_template Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/data_construction/seed_to_vignette.py
code served (permissive licence) · get_code("21e4de4c0eededee")
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)

# Define the arguments for build_seed2vignette_template
ARGS = {
    "data_type": "financial account passwords and PIN numbers",
    "data_subject": "Sarah Chen",
    "data_sender": "Michael, a financial advisor",
    "data_recipient": "corporate compliance officer",
    "transmission_principle": "send an email"
}
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
load_json Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/helper/inspect_data.py
code served (permissive licence) · get_code("c56cba76d2b331d1")
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 io

# Create a file-like object containing valid JSON
json_content = json.dumps({"key": "value", "number": 42, "list": [1, 2, 3]})
uploaded_file = io.StringIO(json_content)

ARGS = {
    "uploaded_file": uploaded_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
post_process Ran
ran; our contract draft was wrong, not the code
SALT-NLP/PrivacyLens/evaluation/get_final_action.py
code served (permissive licence) · get_code("bc753da2fbd7dbb7")
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) 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)

# The function post_process takes a single string parameter 's'
# It performs various string operations: startswith, slicing, find, strip
# No complex objects needed - just a string

ARGS = {
    's': '<|start_header_id|>assistant<|end_header_id|>This is a test response with some content} and more text'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
prepare_prompt_to_evaluate_helpfulness Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/evaluation/evaluate_final_action.py
code served (permissive licence) · get_code("0e201f4f96ed959a")
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)

ARGS = {
    "user_name": "Alice Johnson",
    "user_instruction": "Send an email to bob@example.com with the quarterly report attached",
    "trajectory": "1. Agent opened email application\n2. Agent clicked compose button\n3. Agent entered recipient: bob@example.com\n4. Agent selected file: quarterly_report.pdf",
    "action": "Agent clicked send button to send email with quarterly_report.pdf to bob@example.com"
}
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
prepare_prompt_to_identify_leakage Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/evaluation/evaluate_final_action.py
code served (permissive licence) · get_code("923a15549e1d97f4")
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

# The function prepare_prompt_to_identify_leakage takes three string parameters:
# name, secret, and action. It constructs and returns a prompt string.
# No complex objects are needed - just strings.

ARGS = {
    "name": "John Smith",
    "secret": "The patient is suffering from severe anxiety disorder and takes medication X daily.",
    "action": "EmailSendAction to recipient@example.com with subject 'Patient Update' containing details about the patient's anxiety condition and medication.",
}
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
prepare_prompt_to_identify_sensitive_information Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/evaluation/evaluate_final_action.py
code served (permissive licence) · get_code("4b16c77d829248d2")
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

# Analyze the function: it takes data_type, data_subject, and trajectory strings
# and constructs a prompt for identifying sensitive information.
# All parameters are strings used in f-string formatting.

ARGS = {
    "data_type": "financial transaction record",
    "data_subject": "the account holder",
    "trajectory": """Action: DatabaseQuery
Action Input: {"table": "accounts", "filter": "account_id=5678"}

Observation: {"results": [{"account_id": "5678", "holder_name": "Jane Patterson", "balance": "$250,000", "recent_transactions": [{"date": "2024-01-15", "amount": "$5,000", "description": "Transfer to savings"}, {"date": "2024-01-10", "amount": "$12,000", "description": "Wire transfer to external bank"}]}]}

Action: NotionManagerSearchContent
Action Input: {"keywords": "tax documents"}

Observation: {"results": [{"id": "11111", "title": "Tax Documentation Notes", "content": "Jane Patterson's tax filing shows business income of $180,000 and deductions of $45,000. Previous years show similar patterns of income."}]}"""
}
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
prepare_seed_probing_prompt Ran
ran; our contract draft was wrong, not the code
SALT-NLP/PrivacyLens/evaluation/probing.py
code served (permissive licence) · get_code("cefb3c272dc7c713")
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) 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 = {
    "data_type": "a personal email address",
    "data_subject": "Alice",
    "data_sender": "bob@example.com",
    "data_sender_name": "Bob",
    "data_recipient": "charlie@example.com",
    "transmission_principle": "share via email"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
prepare_vignette_probing_prompt Ran
ran; our contract draft was wrong, not the code
SALT-NLP/PrivacyLens/evaluation/probing.py
code served (permissive licence) · get_code("f3c42f0c80510056")
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) 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 = {
    'data_type': 'a medical record',
    'data_subject': 'Alice',
    'data_sender': 'Dr. Smith',
    'data_recipient': 'Insurance Company',
    'transmission_principle': 'share via email',
    'story': 'Alice visited her doctor for a routine checkup. Dr. Smith documented her medical history and test results.'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
print_api_usage Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/helper/utils.py
code served (permissive licence) · get_code("9725802be067e765")
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 threading

class APIUsageTracker(object):
    """A singleton class to track API usage."""
    _instance = None
    _token_usage_lock = threading.Lock()

    def __new__(cls):
        if cls._instance is None:
            cls._instance = super(APIUsageTracker, cls).__new__(cls)
            cls._instance.token_usage = {}
        return cls._instance

    def get_token_usage(self):
        return self.token_usage

    def reset_token_usage(self):
        self.token_usage = {}

    def increment_token_usage(self, model, prompt_tokens, completion_tokens):
        with self._token_usage_lock:
            if model in self.token_usage:
                self.token_usage[model]['prompt_tokens'] += prompt_tokens
                self.token_usage[model]['completion_tokens'] += completion_tokens
            else:
                self.token_usage[model] = {'prompt_tokens': prompt_tokens, 'completion_tokens': completion_tokens}

api_usage_tracker = APIUsageTracker()

def print_api_usage(func):
    """A decorator to print API usage even if an exception is raised."""

    def wrapper(*args, **kwargs):
        try:
            func(*args, **kwargs)
        except Exception as e:
            print(f'{func.__name__} raised an exception: {e}')
            raise
        finally:
            print('API Usage Information:')
            print(api_usage_tracker.get_token_usage())
    return wrapper


def sample_function():
    """A simple function to be decorated."""
    api_usage_tracker.increment_token_usage('gpt-3.5', 100, 50)
    pass


decorated_func = print_api_usage(sample_function)

ARGS = {
    'func': sample_function
}
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 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
process_trajectory Ran
ran; our contract draft was wrong, not the code
SALT-NLP/PrivacyLens/data_construction/format_trajectory.py
code served (permissive licence) · get_code("41bfcf1b7980d8c2")
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) and returned list, 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 torch

# Construct a valid case dict for process_trajectory
case = {
    'intermediate_steps': [
        # Valid step: [action_tuple, observation_tuple]
        [
            ('ActionType', 'param1'),
            ('observation_result',)
        ],
        # Another valid step
        [
            ('AnotherAction', 'param2'),
            ('another_observation',)
        ],
        # Step with a blocked action type - will cause break
        [
            ('GmailSendEmail', 'param3'),
            ('blocked_observation',)
        ]
    ]
}

ARGS = {
    'case': case
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, 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
retry Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/helper/utils.py
code served (permissive licence) · get_code("28fb684d6e720c67")
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 functools
import random
import time

def retry(max_retries=5, initial_delay=1, backoff_factor=2, exceptions=(Exception,), jitter=False):
    """
    A universal retry decorator with increasing delay.

    Parameters:
    - max_retries (int): The maximum number of retries.
    - initial_delay (int or float): The initial delay between retries in seconds.
    - backoff_factor (int or float): The factor by which the delay is multiplied in each retry.
    - exceptions (tuple): A tuple of exception classes to catch and retry. Defaults to (Exception,), which catches all exceptions.
    - jitter (bool): If True, adds a small random amount to the delay to avoid thundering herd problem.
    """

    def decorator(func):

        @functools.wraps(func)
        def wrapper(*args, **kwargs):
            delay = initial_delay
            for attempt in range(max_retries + 1):
                try:
                    return func(*args, **kwargs)
                except exceptions as e:
                    if attempt < max_retries:
                        total_delay = delay + random.uniform(0, delay * 0.1) if jitter else delay
                        print(f'Retry {attempt + 1} of {max_retries} after error: {e}. Waiting {total_delay} seconds...')
                        time.sleep(total_delay)
                        delay *= backoff_factor
                    else:
                        raise
        return wrapper
    return decorator


# Define a test function to decorate
def sample_function():
    return "success"


# Build ARGS dict with valid parameter values
ARGS = {
    "max_retries": 3,
    "initial_delay": 0.1,
    "backoff_factor": 2,
    "exceptions": (ValueError, RuntimeError),
    "jitter": True
}
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 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
test_no_explicit_word Ran
ran; our contract draft was wrong, not the code
salt-nlp/privacylens/data_construction/seed_to_vignette.py
code served (permissive licence) · get_code("ca248793a28dcb8b")
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 torch
import torch.nn as nn

# Analyze the function:
# - instruction: a string parameter (not used in the function body)
# - output: a string parameter that is split by '\n' and the first line is checked for certain words

# Create test cases
instruction = "Generate a story about something"
output = "This is a normal story about a character.\nMore details here."

ARGS = {
    "instruction": instruction,
    "output": output,
}
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 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

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 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 13 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

As language models (LMs) are widely utilized in personalized communication scenarios (e.g., sending emails, writing social media posts) and endowed with a certain level of agency, ensuring they act in accordance with the contextual privacy norms becomes increasingly critical. However, quantifying the privacy norm awareness of LMs and the emerging privacy risk in LM-mediated communication is challenging due to (1) the contextual and long-tailed nature of privacy-sensitive cases, and (2) the lack of evaluation approaches that capture realistic application scenarios. To address these challenges, we propose PrivacyLens, a novel framework designed to extend privacy-sensitive seeds into expressive vignettes and further into agent trajectories, enabling multi-level evaluation of privacy leakage in LM agents' actions. We instantiate PrivacyLens with a collection of privacy norms grounded in privacy literature and crowdsourced seeds. Using this dataset, we reveal a discrepancy between LM performance in answering probing questions and their actual behavior when executing user instructions in an agent setup. State-ofthe-art LMs, like GPT-4 and Llama-3-70B, leak sensitive information in 25.68% and 38.69% of cases, even when prompted with privacy-enhancing instructions. We also demonstrate the dynamic nature of PrivacyLens by extending each seed into multiple trajectories to red-team LM privacy leakage risk. Dataset and code are available at https://github.com/SALT-NLP/PrivacyLens. * The work started when the author was a postdoc at Stanford. 38th Conference on Neural Information Processing Systems (NeurIPS 2024) Track on Datasets and Benchmarks.

For agents

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

get_harvested_code_for_paper("2409.00138")
get_code_for_paper("2409.00138")
have("2409.00138")

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

curl https://syntology.ai/api/ran/2409.00138.json

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

[![Syntology run record](https://syntology.ai/api/ran/2409.00138.svg)](https://syntology.ai/paper/2409.00138)

Connect an agent — have() is free.