SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.17419 · EMNLP · 2024

Leave No Document Behind: Benchmarking Long-Context LLMs with Extended Multi-Doc QA

Lei Zhang, Haiyang Yu, Fei Huang, Yongbin Li, Nan Xu, Cheng Fu, Xinghua Zhang, Minzheng Wang, Bingli Wu, Min Yang, Shengyi Liao, Longze Chen, and 2 more

arXiv · PDF · Open in the Atlas

Code that ran

6 of 8 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; 2 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 4 failures of our instrument; 2 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/2406.17419.json · badge: /api/ran/2406.17419.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
MozerWang/Loong canonical (classifier label)
classifier basis: census
6 ran · 2 no recorded run
FunctionStatusWhere it lives
continue_gen Ran
ran; our contract draft was wrong, not the code
MozerWang/Loong/src/utils/util.py
code served (permissive licence) · get_code("0f267f51086f55f7")
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 json
import tempfile
import os

# Create a temporary file for input_path
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.jsonl')
temp_file_path = temp_file.name

# Write some initial data to the file
initial_data = [
    {'id': '1', 'text': 'example1', 'processed': True},
    {'id': '2', 'text': 'example2', 'processed': False},
    {'id': '3', 'text': 'example3', 'processed': True},
]

for item in initial_data:
    temp_file.write(json.dumps(item, ensure_ascii=False) + '\n')
temp_file.close()

# Create gen_data with some items that exist in the file and some that don't
gen_data = [
    {'id': '1', 'text': 'new_example1', 'processed': False},
    {'id': '4', 'text': 'new_example4', 'processed': True},
    {'id': '2', 'text': 'new_example2', 'processed': False},
    {'id': '5', 'text': 'new_example5', 'processed': True},
]

tag = 'processed'

ARGS = {
    'input_path': temp_file_path,
    'gen_data': gen_data,
    'tag': tag,
}
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
count_lines Ran
ran, honoured the contract we drafted
MozerWang/Loong/src/utils/util.py
code served (permissive licence) · get_code("2de377da8e0f3e4b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import os

# Create a temporary file with some lines of text
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, encoding='utf-8', suffix='.txt')
temp_file.write("Line 1\n")
temp_file.write("Line 2\n")
temp_file.write("Line 3\n")
temp_file.write("Line 4\n")
temp_file.write("Line 5\n")
temp_file.close()

file_path = temp_file.name

ARGS = {
    "file_path": 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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
extract_number Ran
ran, honoured the contract we drafted
MozerWang/Loong/src/utils/metric.py
code served (permissive licence) · get_code("02bbb0545f754a36")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re
import json

def extract_number(text):
    match = re.search('\\[\\[([0-9]*\\.?[0-9]+)\\]\\]', text)
    if match:
        return float(match.group(1))
    match = re.search('\\[([0-9]*\\.?[0-9]+)\\]', text)
    if match:
        return float(match.group(1))
    return None

ARGS = {
    "text": "The value is [[42.5]] in the data"
}
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 float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
failure_prompts Ran
ran; our contract draft was wrong, not the code
MozerWang/Loong/src/utils/metric.py
code served (permissive licence) · get_code("c02d809b6cb29b4c")
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 json
import tempfile
import os

# Create temporary files with the required JSON data
temp_dir = tempfile.mkdtemp()

# Create old_evaluate_output_path file with evaluation results
eval_output_path = os.path.join(temp_dir, "eval_output.jsonl")
with open(eval_output_path, 'w') as f:
    # Line 1: valid sample with score
    f.write(json.dumps({'id': '1', 'tag_score': '[[0.85]]', 'generate_response': 'some response'}) + '\n')
    # Line 2: invalid sample - no score extracted
    f.write(json.dumps({'id': '2', 'tag_score': 'invalid', 'generate_response': 'response'}) + '\n')
    # Line 3: invalid sample - empty generate_response
    f.write(json.dumps({'id': '3', 'tag_score': '[0.75]', 'generate_response': ''}) + '\n')
    # Line 4: valid sample
    f.write(json.dumps({'id': '4', 'tag_score': '[0.92]', 'generate_response': 'another response'}) + '\n')

# Create old_output_path file with generation data
gen_output_path = os.path.join(temp_dir, "gen_output.jsonl")
with open(gen_output_path, 'w') as f:
    f.write(json.dumps({'id': '1', 'prompt': 'What is AI?', 'question': 'Define AI', 'answer': 'Artificial Intelligence'}) + '\n')
    f.write(json.dumps({'id': '2', 'prompt': 'What is ML?', 'question': 'Define ML', 'answer': 'Machine Learning'}) + '\n')
    f.write(json.dumps({'id': '3', 'prompt': 'What is DL?', 'question': 'Define DL', 'answer': 'Deep Learning'}) + '\n')
    f.write(json.dumps({'id': '4', 'prompt': 'What is NLP?', 'question': 'Define NLP', 'answer': 'Natural Language Processing'}) + '\n')

# Create a simple args object with the required attributes
class Args:
    def __init__(self, old_evaluate_output_path, old_output_path):
        self.old_evaluate_output_path = old_evaluate_output_path
        self.old_output_path = old_output_path

args = Args(eval_output_path, gen_output_path)
tag = 'tag_score'

ARGS = {
    'args': args,
    'tag': tag
}
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
get_content Ran
ran; our contract draft was wrong, not the code
MozerWang/Loong/src/utils/prompt.py
code served (permissive licence) · get_code("8581c9fc4e680d08")
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 json
from pathlib import Path
import os

# Create temporary directory structure with test files
tmpdir = tempfile.mkdtemp()

# Create financial document directory and file
financial_dir = Path(tmpdir) / 'financial'
financial_dir.mkdir(exist_ok=True)
financial_file = financial_dir / '2024-test_doc.txt'
financial_file.write_text('Financial content here')

# Create paper document directory and file
paper_dir = Path(tmpdir) / 'paper'
paper_dir.mkdir(exist_ok=True)
paper_file = paper_dir / 'paper_doc.txt'
paper_file.write_text('# Paper Title\nPaper content here')

# Create legal document directory with JSON file
legal_dir = Path(tmpdir) / 'legal'
legal_dir.mkdir(exist_ok=True)
legal_json_path = legal_dir / 'legal.json'
legal_data = {
    'case_001': {
        'content': 'Legal content describing the case',
        'result': 'Legal result and judgment'
    }
}
with open(legal_json_path, 'w') as f:
    json.dump(legal_data, f)

# Create args object
class Args:
    def __init__(self, doc_path):
        self.doc_path = doc_path

args = Args(tmpdir)

# Create item for financial doc (level != '4')
item_financial = {
    'type': 'financial',
    'level': '1',
    'instruction': 'Read financial doc'
}

# Create item for paper doc
item_paper = {
    'type': 'paper',
    'level': '2',
    'instruction': 'Read paper'
}

# Create item for legal doc
item_legal = {
    'type': 'legal',
    'level': '3',
    'instruction': 'Read legal document'
}

# Define ARGS with all parameters
ARGS = {
    'args': args,
    'item': item_legal,
    'doc_name': 'case_001',
    'idx': 0
}
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
get_contents Ran
ran; our contract draft was wrong, not the code
MozerWang/Loong/src/utils/prompt.py
code served (permissive licence) · get_code("eea848762af9fe85")
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 tempfile
import json
from pathlib import Path
import os

# Create temporary directory structure for testing
tmpdir = tempfile.mkdtemp()
doc_path = Path(tmpdir)

# Create financial directory
financial_dir = doc_path / 'financial'
financial_dir.mkdir()

# Create a sample financial document
financial_file = financial_dir / 'report-2024-financial_report.txt'
financial_file.write_text('This is a financial report content.')

# Create another financial document for level 4
financial_file_level4 = financial_dir / 'report-financial_data.txt'
financial_file_level4.write_text('This is financial data content.')

# Create paper directory
paper_dir = doc_path / 'paper'
paper_dir.mkdir()

# Create a sample paper document
paper_file = paper_dir / 'sample_paper.txt'
paper_file.write_text('# Sample Paper Title\nThis is the paper content.')

# Create legal directory with legal.json
legal_dir = doc_path / 'legal'
legal_dir.mkdir()

legal_json_file = legal_dir / 'legal.json'
legal_data = {
    'case_001': {
        'content': 'Case content for case 001',
        'result': 'Judgment result for case 001'
    },
    'case_002': {
        'content': 'Case content for case 002',
        'result': 'Judgment result for case 002'
    }
}
legal_json_file.write_text(json.dumps(legal_data))

# Create args object with doc_path attribute
class Args:
    def __init__(self, doc_path):
        self.doc_path = doc_path

args = Args(str(doc_path))

# Create item for financial document (level != 4)
item_financial = {
    'type': 'financial',
    'level': 1,
    'instruction': 'Process financial report'
}

# Create item for legal document
item_legal = {
    'type': 'legal',
    'level': 1,
    'instruction': 'Process legal case'
}

# Create item for legal document level 4
item_legal_level4 = {
    'type': 'legal',
    'level': 4,
    'instruction': '阅读以上判决文书,我将给你若干份判决结果:'
}

# Create item for paper document
item_paper = {
    'type': 'paper',
    'level': 1,
    'instruction': 'Process paper'
}

# Test doc_names
doc_names_financial = ['financial_report']
doc_names_legal = ['case_001', 'case_002']
doc_names_paper = ['sample_paper.txt']

ARGS = {
    'args': args,
    'item': item_legal,
    'doc_names': doc_names_legal
}
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
cal_metric Not yet run
unverified (harvested, no recorded run)
MozerWang/Loong/src/utils/metric.py
code served (permissive licence) · get_code("08724221f8dab062")
How it ran

no invocation recorded.

get_doc_str Not yet run
unverified (harvested, no recorded run)
MozerWang/Loong/src/utils/prompt.py
code served (permissive licence) · get_code("c62f3f6d440ecf4d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 8 samples lifted from this paper's linked repositories, 8 carry a function identity and 0 do not. Of the 8 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 7 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
count_lines 1 other paper
1 dated earlier
2405.16871 (2024-05)
Multi-Behavior Generative Recommendation

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

Long-context modeling capabilities have garnered widespread attention, leading to the emergence of Large Language Models (LLMs) with ultra-context windows. Meanwhile, benchmarks for evaluating long-context LLMs are gradually catching up. However, existing benchmarks employ irrelevant noise texts to artificially extend the length of test cases, diverging from the real-world scenarios of longcontext applications. To bridge this gap, we propose a novel long-context benchmark, Loong, aligning with realistic scenarios through extended multi-document question answering (QA). Unlike typical document QA, in Loong's test cases, each document is relevant to the final answer, ignoring any document will lead to the failure of the answer. Furthermore, Loong introduces four types of tasks with a range of context lengths: Spotlight Locating, Comparison, Clustering, and Chain of Reasoning, to facilitate a more realistic and comprehensive evaluation of long-context understanding. Extensive experiments indicate that existing longcontext language models still exhibit considerable potential for enhancement. Retrieval augmented generation (RAG) achieves poor performance, demonstrating that Loong can reliably assess the model's long-context modeling capabilities. 1

For agents

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

get_harvested_code_for_paper("2406.17419")
get_code_for_paper("2406.17419")
have("2406.17419")

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

curl https://syntology.ai/api/ran/2406.17419.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.17419.svg)](https://syntology.ai/paper/2406.17419)

Connect an agent — have() is free.