SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.07539 · NeurIPS · 2025

XIFBench: Evaluating Large Language Models on Multilingual Instruction Following

Juntao Li, Min Zhang, Kehai Chen, Xuefeng Bai, Zhenyu Li, Yunfei Long, Yaoyin Zhang, Xuchen Wei, En Zh, Constr Category, Constr Content

arXiv · PDF · Open in the Atlas

Code that ran

11 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 10 have no contract verdict; 1 failure of our instrument; 1 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/2503.07539.json · badge: /api/ran/2503.07539.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
zhenyuli801/XIFBench canonical (classifier label)
classifier basis not recorded on the link
11 ran · 1 no recorded run
FunctionStatusWhere it lives
format_evaluation_requirements Ran
ran on a synthesized input
zhenyuli801/XIFBench/D2_validate_translations.py
code served (permissive licence) · get_code("3a43b74a35f656eb")
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 = {
    "evaluation_requirements": [
        {
            "requirement_category": "Accuracy",
            "requirement_direction": "maximize",
            "requirement_specification": "Translation should preserve the original meaning"
        },
        {
            "requirement_category": "Fluency",
            "requirement_direction": "maximize",
            "requirement_specification": "Output should read naturally in the target language"
        },
        {
            "requirement_category": "Latency",
            "requirement_direction": "minimize",
            "requirement_specification": "Translation should complete within 100ms"
        }
    ]
}
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
format_sampled_constraints Ran
ran on a synthesized input
zhenyuli801/XIFBench/B2_combine_constraints.py
code served (permissive licence) · get_code("1cfb172de71361fd")
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)

# Create sample constraint dictionaries
sampled_constraints = [
    {
        "constraint_category": "Safety",
        "constraint_direction": "maximize",
        "constraint_specification": "The response should not contain harmful content"
    },
    {
        "constraint_category": "Relevance",
        "constraint_direction": "minimize",
        "constraint_specification": "Avoid off-topic information in the answer"
    },
    {
        "constraint_category": "Clarity",
        "constraint_direction": "maximize",
        "constraint_specification": "The explanation should be easy to understand"
    }
]

ARGS = {
    "sampled_constraints": sampled_constraints
}
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
generate_reasoning_and_responses Ran
ran on a synthesized input
zhenyuli801/XIFBench/G1_api_inference.py
code served (permissive licence) · get_code("635301a89612a3d4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 2, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import time
from tqdm import tqdm
from loguru import logger
from openai import OpenAI
from unittest.mock import Mock, patch, MagicMock

# Create mock response objects that mimic OpenAI API responses
mock_message = Mock()
mock_message.reasoning_content = "This is the reasoning for the response."
mock_message.content = "This is the generated response."

mock_choice = Mock()
mock_choice.message = mock_message

mock_response = Mock()
mock_response.choices = [mock_choice]

# Patch the OpenAI client to return our mock response
def mock_api_call(*args, **kwargs):
    return mock_response

# Create sample prompts
sample_prompts = [
    "What is the capital of France?",
    "Explain quantum computing in simple terms.",
]

# Define the arguments dictionary
ARGS = {
    'prompts': sample_prompts,
    'api_key': 'sk-test-key-12345',
    'base_url': 'https://api.openai.com/v1',
    'model': 'gpt-4-turbo',
    'max_tokens': 1024,
    'temperature': 0.0,
    'max_api_retry': 5,
    'base_wait_time': 1,
    'debug_mode': False,
}

# Apply the mock to make the function executable without actual API calls
with patch('openai.OpenAI') as mock_openai_class:
    mock_client = Mock()
    mock_client.chat.completions.create = mock_api_call
    mock_openai_class.return_value = mock_client
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
loguru, openai, tqdm — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 2, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_alpacaeval Ran
ran on a synthesized input
zhenyuli801/XIFBench/A1_sample_instructions.py
code served (permissive licence) · get_code("1aac49f7aa943d77")
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 JSON file with AlpacaEval-like data
temp_dir = tempfile.mkdtemp()
test_file = os.path.join(temp_dir, "alpacaeval_test.json")

# Create sample AlpacaEval data
sample_data = [
    {
        "instruction": "What is the capital of France?",
        "input": "",
        "output": "The capital of France is Paris."
    },
    {
        "instruction": "Write a short poem about nature.",
        "input": "",
        "output": "Leaves dance in the wind,\nBirds sing their sweet song,\nNature's beauty endless."
    }
]

# Write to temporary file
with open(test_file, 'w', encoding='utf-8') as f:
    json.dump(sample_data, f)

ARGS = {
    "file_path": test_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 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
load_json Ran
ran; our contract draft was wrong, not the code
zhenyuli801/XIFBench/A1_sample_instructions.py
code served (permissive licence) · get_code("cf254803acaaf54c")
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.json")

sample_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "data"}
}

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

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_jsonl Ran
ran on a synthesized input
zhenyuli801/XIFBench/A1_sample_instructions.py
code served (permissive licence) · get_code("d4c04971c3f6dc69")
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 JSONL file for testing
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "test.jsonl")

# Write sample JSONL data
sample_data = [
    {"id": 1, "text": "hello world", "value": 42},
    {"id": 2, "text": "foo bar", "value": 3.14},
    {"id": 3, "text": "test data", "value": True}
]

with open(temp_file, 'w', encoding='utf-8') as f:
    for item in sample_data:
        f.write(json.dumps(item) + '\n')

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 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
parse_constraint_brainstorming_response Ran
ran on a synthesized input
zhenyuli801/XIFBench/B1_brainstorm_constraints.py
code served (permissive licence) · get_code("c104e11d10734d8e")
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 re
from loguru import logger

ARGS = {
    "response": """# Content constraints
- **clarity**: The constraint should be easy to understand and interpret
- **specificity**: The constraint must be precise and well-defined

# Format constraints
- **conciseness**: Keep constraint descriptions brief and to the point
- **structure**: Use consistent formatting across all constraints

# Relevance constraints
- **applicability**: Ensure constraints are relevant to the problem domain"""
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
loguru — 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
parse_constraint_combination_response Ran
ran on a synthesized input
zhenyuli801/XIFBench/B2_combine_constraints.py
code served (permissive licence) · get_code("e02d2d2f5e843460")
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 = {
    "response": "  This is a combined instruction from constraint combination.  "
}
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
parse_requirement_categorization_response Ran
ran on a synthesized input
zhenyuli801/XIFBench/C2_categorize_requirements.py
code served (permissive licence) · get_code("505083d5b2dec3e6")
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 re
from loguru import logger

VALID_REQUIREMENT_CATEGORIES = {'content', 'style', 'situation', 'format', 'numerical'}
INVALID_CATEGORY_TAG = '[INVALID_CATEGORY]'

def parse_requirement_categorization_response(response: str) -> list[dict[str, str]]:
    """ Parse the response from Requirement Categorization and extract the requirement categories. """
    parsed_data = []
    CATEGORY_DIRECTION_PATTERN = '^\\d+\\.\\s*(.+?)\\s*-\\s*(.+)'
    try:
        lines = response.strip().splitlines()
        for line in lines:
            line = line.strip()
            if not line:
                continue
            category_match = re.match(CATEGORY_DIRECTION_PATTERN, line)
            if category_match:
                category = category_match.group(1).lower()
                direction = category_match.group(2).lower()
                category = category.strip('*')
                direction = direction.strip('*')
                category = category.split(' ')[0]
                if category in VALID_REQUIREMENT_CATEGORIES:
                    parsed_data.append({'requirement_category': category, 'requirement_direction': direction})
                else:
                    logger.warning(f'Unexpected category encountered: {category}')
                    parsed_data.append({'requirement_category': f'{INVALID_CATEGORY_TAG} - {category}', 'requirement_direction': f'{INVALID_CATEGORY_TAG} - {direction}'})
        return parsed_data
    except Exception as e:
        logger.error(f'Failed to extract requirement categories from response. Error: {e}')
        return None

# Setup code
response_valid = """1. content - increase clarity and precision
2. style - use active voice
3. format - organize in bullet points
4. numerical - provide exact metrics"""

response_with_invalid = """1. content - increase clarity
2. invalid_cat - some direction
3. style - use passive voice"""

response_empty = ""

response_malformed = """This is not formatted correctly
No numbered items here"""

ARGS = {
    'response': response_valid
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
loguru — 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
parse_requirement_decomposition_response Ran
ran on a synthesized input
zhenyuli801/XIFBench/C1_decompose_requirements.py
code served (permissive licence) · get_code("d8bad0118c48538d")
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 re
from loguru import logger

def parse_requirement_decomposition_response(response: str) -> list[str]:
    """ Parse the response from Requirement Decomposition and extract the evaluation requirements. """
    requirements = []
    REQUIREMENT_PATTERN = '^\\d+\\.\\s*(.+)'
    try:
        lines = response.strip().splitlines()
        for line in lines:
            line = line.strip()
            if not line:
                continue
            requirement_match = re.match(REQUIREMENT_PATTERN, line)
            if requirement_match:
                requirement = requirement_match.group(1).strip()
                requirements.append(requirement)
        return requirements
    except Exception as e:
        logger.error(f'Failed to extract requirements from response. Error: {e}')
        return None


# Setup code to construct valid arguments
response_text = """1. The system shall validate user input before processing
2. The system shall return results within 2 seconds
3. The system shall handle edge cases gracefully

4. Additional requirement with proper formatting"""

ARGS = {
    "response": response_text
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
loguru — 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
sample_constraints Ran
ran on a synthesized input
zhenyuli801/XIFBench/B2_combine_constraints.py
code served (permissive licence) · get_code("33730eed64c4e455")
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. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import numpy as np

# Create brainstormed constraints with the expected structure
brainstormed_constraints = [
    {
        'constraints_category': 'grammar',
        'constraints_content': [
            {'rule': 'subject-verb agreement', 'description': 'Ensure subject and verb agree in number'},
            {'rule': 'tense consistency', 'description': 'Maintain consistent verb tense throughout'},
        ]
    },
    {
        'constraints_category': 'style',
        'constraints_content': [
            {'rule': 'formal tone', 'description': 'Use formal language'},
            {'rule': 'conciseness', 'description': 'Keep sentences short and clear'},
        ]
    },
    {
        'constraints_category': 'semantic',
        'constraints_content': [
            {'rule': 'factual accuracy', 'description': 'Ensure all statements are factually correct'},
            {'rule': 'logical flow', 'description': 'Maintain logical progression of ideas'},
        ]
    },
]

max_constraint_count = 6
constraint_sample_times = 3
random_seed = 42

ARGS = {
    'brainstormed_constraints': brainstormed_constraints,
    'max_constraint_count': max_constraint_count,
    'constraint_sample_times': constraint_sample_times,
    'random_seed': random_seed,
}
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 — a list 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
get_openai_response Not yet run
unverified (harvested, no recorded run)
zhenyuli801/XIFBench/B1_brainstorm_constraints.py
code served (permissive licence) · get_code("d52e4d9cbd07b9c1")
How it ran

no invocation recorded.

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
load_json 1 other paper
1 later
2606.10403 (2026-06)
KCSAT-ML: Probing Reasoning Models with Nationwide-Cohort Human Difficulty

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) have demonstrated remarkable instructionfollowing capabilities across various applications. However, their performance in multilingual settings lacks systematic investigation, with existing evaluations lacking fine-grained constraint analysis across diverse linguistic contexts. We introduce XIFBench, a comprehensive constraint-based benchmark for evaluating multilingual instruction-following abilities of LLMs, comprising 558 instructions with 0-5 additional constraints across five categories (Content, Style, Situation, Format, and Numerical) in six languages spanning different resource levels. To support reliable and consistent cross-lingual evaluation, we implement three methodological innovations: cultural accessibility annotation, constraint-level translation validation, and requirement-based evaluation using English requirements as semantic anchors across languages. Extensive experiments with various LLMs not only quantify performance disparities across resource levels but also provide detailed insights into how language resources, constraint categories, instruction complexity, and cultural specificity influence multilingual instruction-following. Our code and data are available at https://github.com/zhenyuli801/XIFBench.

For agents

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

get_harvested_code_for_paper("2503.07539")
get_code_for_paper("2503.07539")
have("2503.07539")

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

curl https://syntology.ai/api/ran/2503.07539.json

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

[![Syntology run record](https://syntology.ai/api/ran/2503.07539.svg)](https://syntology.ai/paper/2503.07539)

Connect an agent — have() is free.