SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.15028 · 2026

Long Context, Less Focus: A Scaling Gap in LLMs Revealed through Privacy and Personalization

Shangding Gu

arXiv · PDF · Open in the Atlas

Code that ran

5 of 5 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; 5 failures of our instrument; 0 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/2602.15028.json · badge: /api/ran/2602.15028.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
SafeRL-Lab/PAPerBench canonical (classifier label)
classifier basis: census
5 ran · 0 no recorded run
FunctionStatusWhere it lives
build_user Ran
ran; our contract draft was wrong, not the code
SafeRL-Lab/PAPerBench/paperbench/personalization/eval_personalization_mcq.py
code served (permissive licence) · get_code("6a4b1aa03938d905")
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
from typing import Dict, Any

# Construct a valid sample dictionary matching what build_user expects
sample = {
    'specific_generated_context': 'The capital of France is Paris. It is known for the Eiffel Tower.',
    'generated_query': 'What is the most famous landmark in Paris?',
    'mcq': {
        'options': {
            'A': {'text': 'The Eiffel Tower'},
            'B': {'text': 'The Louvre Museum'},
            'C': {'text': 'Notre-Dame Cathedral'},
            'D': {'text': 'Arc de Triomphe'},
            'E': {'text': 'Sacré-Cœur'}
        }
    }
}

ARGS = {
    'sample': sample
}
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
extract_letter Ran
ran; our contract draft was wrong, not the code
SafeRL-Lab/PAPerBench/paperbench/personalization/eva_personalization_litellm.py
code served (permissive licence) · get_code("792893ef9eb73f12")
How it ran

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

import re

LETTER_RE = re.compile('\\b([ABCDE])\\b', re.IGNORECASE)

def extract_letter(text: str) -> str:
    if not text:
        return ''
    m = LETTER_RE.search(text.strip())
    return m.group(1).upper() if m else ''

ARGS = {
    'text': 'The letter A is here'
}
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
extract_letter Ran
ran; our contract draft was wrong, not the code
SafeRL-Lab/PAPerBench/paperbench/personalization/eval_personalization_mcq.py
code served (permissive licence) · get_code("41dae4f2e4468ab8")
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: extract_letter takes a single parameter 'text' of type str
# The function searches for letters A-E in the text and returns the first match in uppercase
# If no match is found, it returns an empty string

ARGS = {
    'text': 'The answer is B for sure'
}
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
normalize_options_to_AE Ran
ran; our contract draft was wrong, not the code
SafeRL-Lab/PAPerBench/paperbench/personalization/eva_personalization_litellm.py
code served (permissive licence) · get_code("4fac52adc0075074")
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
from typing import Any, Dict, Optional, List

torch.manual_seed(0)

# Test case 1: dict with A-E keys
opts_dict_ae = {
    "A": "option A text",
    "B": "option B text",
    "C": "option C text",
    "D": "option D text",
    "E": "option E text"
}

# Test case 2: dict with lowercase a-e keys
opts_dict_lowercase = {
    "a": "option a text",
    "b": "option b text",
    "c": "option c text",
    "d": "option d text",
    "e": "option e text"
}

# Test case 3: dict with nested 'options' key
opts_dict_nested = {
    "options": {
        "A": "option A",
        "B": "option B",
        "C": "option C",
        "D": "option D",
        "E": "option E"
    }
}

# Test case 4: list with label keys
opts_list_label = [
    {"label": "A", "text": "option A"},
    {"label": "B", "text": "option B"},
    {"label": "C", "text": "option C"},
    {"label": "D", "text": "option D"},
    {"label": "E", "text": "option E"}
]

# Test case 5: list with option keys
opts_list_option = [
    {"option": "A", "prompt": "prompt A"},
    {"option": "B", "prompt": "prompt B"},
    {"option": "C", "prompt": "prompt C"},
    {"option": "D", "prompt": "prompt D"},
    {"option": "E", "prompt": "prompt E"}
]

# Test case 6: None
opts_none = None

# Use the first test case as the main argument
ARGS = {
    "opts_any": opts_dict_ae
}
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
opt_text Ran
ran; our contract draft was wrong, not the code
SafeRL-Lab/PAPerBench/paperbench/personalization/eval_personalization_mcq.py
code served (permissive licence) · get_code("41e469ca57dc7271")
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 torch.nn as nn
from typing import Dict, Any, List

# opt_text function accepts:
# - options: Dict[str, Any] - a dictionary that may contain string keys mapping to values
# - k: str - a key to look up in the options dict
# 
# The function calls:
# - options.get(k, {}) - retrieves value or default empty dict
# - isinstance(v, dict) - checks if value is a dict
# - v.get('text', '') - if dict, gets 'text' key with default ''
# - str(v).strip() - converts to string and strips whitespace

# Test case 1: options has k as a dict with 'text' key
options_dict_with_text = {
    'config': {'text': '  hello world  '},
    'other': 'value'
}

# Test case 2: options has k as a dict without 'text' key
options_dict_no_text = {
    'setting': {'value': 123},
}

# Test case 3: options has k as a string
options_string = {
    'name': '  test string  '
}

# Test case 4: options missing the key entirely
options_empty = {
    'other_key': 'other_value'
}

# Use the first test case for ARGS
ARGS = {
    'options': {
        'config': {'text': '  hello world  '},
        'model': {'type': 'bert', 'text': '  sample text  '},
        'description': 'some value'
    },
    'k': 'config'
}
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

Also present in other papers' code

Of 5 samples lifted from this paper's linked repositories, 5 carry a function identity and 0 do not. Of the 4 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 4 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 deployed in privacy-critical and personalizationoriented scenarios, yet the role of context length in shaping privacy leakage and personalization effectiveness remains largely unexplored. We introduce a large-scale benchmark, PAPerBench, to systematically study how increasing context length influences both personalization quality and privacy protection in LLMs. The benchmark comprises approximately 29,000 instances with context lengths ranging from 1K to 256K tokens, yielding a total of 377K evaluation questions. It jointly evaluates personalization performance and privacy risks across diverse scenarios, enabling controlled analysis of long-context model behavior. Extensive evaluations across state-of-the-art LLMs reveal consistent performance degradation in both personalization and privacy as context length increases. We further provide a theoretical analysis of attention dilution under context scaling, explaining this behavior as an inherent limitation of soft attention in fixed-capacity Transformers. The empirical and theoretical findings together suggest a general scaling gap in current models-long context, less focus. We release the benchmark to support reproducible evaluation and future research on scalable privacy and personalization. Code and data are available at https://github.com/ SafeRL-Lab/PAPerBench.

For agents

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

get_harvested_code_for_paper("2602.15028")
get_code_for_paper("2602.15028")
have("2602.15028")

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

curl https://syntology.ai/api/ran/2602.15028.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.15028.svg)](https://syntology.ai/paper/2602.15028)

Connect an agent — have() is free.