DataFlow

DataFlow

📰 0. News

• [2026-09-12] 🎉 DataFlow has been accepted by ACM SIGMOD 2027! Our paper, DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI, has been accepted by ACM SIGMOD 2027.

• [2026-07-18] DataFlow-Harness Released! 🔥DataFlow-Harness lets coding agents build DataFlow pipelines for you. Download and installation entry points are available in DataFlow-WebUI, where you can use DataFlow through either the visual interface or MCP.

• [2026-05-22] DataFlow-Skills is now available! A collection of skills and tutorials for working with DataFlow. 👉 DataFlow-Skills

• [2026-02-02] 🖥️ DataFlow WebUI is now available! Launch the visual pipeline builder with a single command: dataflow webui. Build and run DataFlow pipelines through an intuitive web interface. 👉 WebUI Docs

  

• [2026-01-20] 🌟 Awesome Works Using DataFlow is now live! A new section showcasing open-source projects and research built on DataFlow. Contributions are welcome! 👉 Awesome Works

• [2025-12-19] 🎉 Our DataFlow technical report is now available! Read and cite our work on arXiv: https://arxiv.org/abs/2512.16676

• [2025-11-20] 🤖 Introducing New Data Agents for DataFlow! Try them out and follow the tutorial on Bilibili: https://space.bilibili.com/3546929239689711/lists/6761342?type=season

• [2025-06-28] 🎉 DataFlow is officially released! Our data-centric AI system is now public. Stay tuned for future updates.

🔍 1. What is DataFlow？

DataFlow is a data preparation and training system designed to generate, refine, evaluate, and filter high-quality data for AI from noisy sources (PDF, plain-text, low-quality QA), thereby improving the performance of large language models (LLMs) in specific domains through targeted training (Pre-training, Supervised Fine-tuning, RL training) or RAG system, in domains such as healthcare, finance, legal, and academic research.

Through an operator-based design, DataFlow turns the entire data cleaning workflow into a reproducible, reusable, and shareable pipeline, providing core infrastructure for the Data-Centric AI community. Additionally, we develop an intelligent DataFlow-agent capable of dynamically assembling new pipelines by recombining existing or creating new operators on demand.

🔍 2. Key Features

✅2.1 Ready-to-Use Data Synthesis and Cleaning Pipelines

• High-Quality Training Data Generation
  • Text, Math, and Code data generation (see DataFlow-Instruct-10K for results)
  • Data generation via tools like AgenticRAG and Text2SQL
• Structured Data Extraction
  • Large-scale PDF → QA conversion
  • Large-scale book PDF → Visual-QA conversion
• Scientific Data Workflow Management
  • Text2SQL workflow management (Accepted by ICDE 2026)
  • Math data workflows (Accepted by KDD 2026)

⚙️2.2 Flexible Custom Pipeline Orchestration

• 10+ core operators define interaction patterns and design principles
• 100+ pipeline-specific operators available for reuse or reference
• Full support for creating custom operators — plug-and-play, easily packaged and distributed via GitHub or PyPI

🧠2.3 Reproducible, Reusable, and Shareable Data-Centric AI System

• Data governance algorithms are encapsulated as operator pipelines, enabling reproducibility and fair comparison of different data governance strategies (❤️research-friendly)
• Easily reuse swap underlying large models to analyze the relationship between model performance and data quality quickly
• Built on Python and Git ecosystems for easy distribution, management, and traceability of high-quality, user-defined data governance operators and pipelines (❤️enterprise-friendly)

🛠️ 3. DataFlow Suite

The DataFlow Suite provides the essential infrastructure to automate and scale LLM data preparation with DataFlow main repository. It comprises four tightly integrated layers:

• DataFlow-Skills – A collection of skills and tutorials for working with DataFlow, covering operator development, pipeline construction, and best practices for data-centric AI.

• DataFlow-WebUI – A visual and AI Agent assisted workspace for building, managing, and running DataFlow pipelines. It combines drag-and-drop pipeline editing with natural-language interaction, enabling users to construct and execute data workflows more efficiently.

• DataFlow-Ecosystem – A modular distribution layer that standardizes operator registration. It enables domain-specific modules (e.g., DataFlow-MM, DataFlow-AI4S) to contribute extensible libraries under a unified abstraction.

• RayOrch – A high-performance orchestration layer built on Ray, providing distributed compute scheduling and resource management for massive-scale data tasks.

Together, these components form a unified, extensible environment that transforms raw data into model-ready intelligence.

✅ 4. Why use DataFlow?

Data generation and cleaning are crucial for high-quality models, but for both enterprises and individuals, these tasks are often time-consuming, labor-intensive, and costly. DataFlow provides a one-stop solution to tackle these challenges efficiently. DataFlow offers the following advantages:

• Enhanced Support for Data Synthesis Modules – Seamlessly integrates text, code, and math data generation pipeline for high-quality training datasets.
• PyTorch-like Programming Management – Clear Pipeline → Operator → Prompt hierarchical structure for workflow control.
• Principled and Multi-Category Operator Classification – Operators are systematically organized into multiple functional categories such as generation, evaluation, filtering, and refinement, forming a scientifically grounded, multi-dimensional taxonomy that reflects different stages of data preparation and enables precise operator selection and composition.
• User-Friendly Design for Easy Debugging and Onboarding – Simplified workflow patterns that reduce the learning curve and accelerate experimentation.

🌟 5. Awesome DataFlow

This section highlights open-source projects, research works, and applications built with DataFlow or deeply integrated with the DataFlow ecosystem.

📦 Community Extensions

• dataflow-speech-quality

  • Speech Quality · Data Cleaning · De-identification · Annotation
  • A DataFlow extension for filtering, cleaning, de-identifying, and annotating customer-service speech transcripts.
  • Maintainer: Jiapeng Mi · China Unicom Software Research Institute

• dataflow-pcb-inspection

  • PCB Data · Data Cleaning · Annotation · Synthetic Data
  • A DataFlow extension for organizing PCB defect data, converting existing pixel bounding boxes to Label Studio tasks, supporting multi-image model pre-annotation, review-result merging, and template-based sample synthesis.
  • Maintainer: Pyromind Dynamics

• dataflow-embodied

  • Embodied AI · Robotics · Data Cleaning · LeRobot
  • A DataFlow extension for organizing robot-collected step labels and time ranges, planning cleaning intervals, and exporting validated LeRobot v2.1 datasets from supported sources.
  • Maintainer: Pyromind Dynamics

Have a project built with DataFlow? We welcome community contributions via Pull Requests. See the DataFlow Ecosystem Guide for guidance on creating a DataFlow extension.

🔧 6. How do operators work？

DataFlow operators are designed with simplicity and clarity in mind.

Operators take structured inputs (JSON, JSONL, CSV) and produce high-quality outputs after intelligent processing. Each operator encapsulates a specific data processing task, providing a clean and consistent API that is easy to understand and integrate. The PyTorch-like design makes them intuitive and ready to use, allowing you to quickly build, combine, and customize pipelines without dealing with complex boilerplate code.

For more details, refer to the Operator Documentation. Below is a minimal example demonstrating how to invoke the PromptedGenerator operator:

Example input data (json/jsonl-style):

// input.json
[
  {"problem": "What is 17 + 25?"},
  {"problem": "If x = 3, compute 2x^2 + 1."}
]

Operator invocation code:

from dataflow.operators.core_text import PromptedGenerator
from dataflow.utils.storage import FileStorage
from dataflow.serving import APILLMServing_request

# set input file to global storage class
storage = FileStorage(first_entry_file_name="./input.json",)

# configure LLM serving (e.g., OpenAI API)
# api key needs to be set via `export DF_API_KEY=sk-xxx`
llm_serving = APILLMServing_request(
    api_url="https://api.openai.com/v1/chat/completions",
)

prompted_generator = PromptedGenerator(
    llm_serving=llm_serving,  # pre-configured LLM backend
    system_prompt="Please solve this math problem."
)

prompted_generator.run(
    storage=self.storage.step(),  # data management (details omitted)
    input_key="problem",          # read from this column
    output_key="solution"         # write to this column
)

After running, the operator will append the generated results into output_key. For example, the output data (json/jsonl-style) becomes:

// dataflow_step1.json
[
    {"problem":"What is 17 + 25?","solution":"42"},
    {"problem":"If x = 3, compute 2x^2 + 1.","solution":"19"}
]

🛠️ 7. Pipelines (Click to expand)

🔧 7.1 Ready-to-Use PipeLines

Current Pipelines in Dataflow are as follows:

• 📝 Text Pipeline: Mine question-answer pairs from large-scale plain-text data (mostly crawed from InterNet) for use in SFT and RL training.
  • 
  • [HuggingFace🤗 demo input & output for Text Pipeline]
• 🧠 Reasoning Pipeline: Enhances existing question–answer pairs with (1) extended chain-of-thought, (2) category classification, and (3) difficulty estimation.
  • 
  • [HuggingFace🤗 demo input & output for Reasoning Pipeline]
• 🗃️ Text2SQL Pipeline: Translates natural language questions into SQL queries, supplemented with explanations, chain-of-thought reasoning, and contextual schema information.
  • 
  • [HuggingFace🤗 demo input & output for Text2SQL Pipeline]
• 📚 Knowlege Base Cleaning Pipeline: Extract and structure knowledge from unorganized sources like tables, PDFs, and Word documents into usable entries for downstream RAG or QA pair generation.
  • 
• 🤖 Agentic RAG Pipeline: Identify and extract QA pairs from existing QA datasets or knowledge bases that require external knowledge to answer, for use in downstream training of Agnetic RAG tasks.
  • 

⚙️ 7.2 Flexible Operator PipeLines

In this framework, operators are categorized into Fundamental Operators, Generic Operators, Domain-Specific Operators, and Evaluation Operators, etc., supporting data processing and evaluation functionalities. Please refer to the documentation for details.

🤖 7.3 Agent Guided Pipelines

• DataFlow Agent: An intelligent assistant that performs data analysis, writes custom operators, and automatically orchestrates them into pipelines based on specific task objectives.

  • 
  • [HuggingFace🤗 demo input & output for DataFlow Agent]

⚡ 8. Quick Start

🛠️ 8.1 Environment Setup and Installation

DataFlow supports Python>=3.10 environments, tested passed on Windows, Linux, and MacOS with Python 3.10, 3.11, and 3.12.

Please use the following commands for environment setup and installation👇

We recommend use uv to install DataFlow for speed up.

pip install uv
uv pip install open-dataflow

If you want to use your own GPU for local inference, please use:

pip install uv
uv pip install "open-dataflow[vllm]"

After installation, you can use the following command to check if dataflow has been installed correctly:

dataflow -v

If installed correctly, you should see:

open-dataflow codebase version: 1.0.0
        Checking for updates...
        Local version:  1.0.0
        PyPI newest version:  1.0.0
You are using the latest version: 1.0.0.

🐳 8.2 Docker Installation (Alternative)

We also provide a Dockerfile for easy deployment and a pre-built Docker image for immediate use.

Option 1: Use Pre-built Docker Image

You can directly pull and use our pre-built Docker image:

# Pull the pre-built image
docker pull molyheci/dataflow:cu124

# Run the container with GPU support
docker run --gpus all -it molyheci/dataflow:cu124

# Inside the container, verify installation
dataflow -v

Option 2: Build from Dockerfile

Alternatively, you can build the Docker image from the provided Dockerfile:

# Clone the repository (HTTPS)
git clone https://github.com/OpenDCAI/DataFlow.git
# Or use SSH
# git clone [email protected]:OpenDCAI/DataFlow.git

cd DataFlow

# Build the Docker image
docker build -t dataflow:custom .

# Run the container
docker run --gpus all -it dataflow:custom

# Inside the container, verify installation
dataflow -v

Note: The Docker image includes CUDA 12.4.1 support and comes with vLLM pre-installed for GPU acceleration. Make sure you have NVIDIA Container Toolkit installed to use GPU features.

🚀 8.3 Quick Start with Google Colab

You can start your first DataFlow translation project directly on Google Colab. By following the provided guidelines, you can seamlessly scale from a simple translation example to more complex DataFlow pipelines.

👉 Start DataFlow with Google Colab

📖 8.4 Reference Project Documentation

For detailed usage instructions and getting started guide, please visit our DataFlow Documentation.

🖥️ 8.5 DataFlow-WebUI

DataFlow provides a Web-based UI (WebUI) for visual pipeline construction and execution.

To start DataFlow-WebUI, simply run following command after install the DataFlow main repository:

dataflow webui

This will automatically download and launch the latest DataFlow-WebUI and open it in your browser (http://localhost:8000/ if it does not open automatically).

📚 8.5.1 WebUI Documentation

• Chinese: DataFlow-WebUI Documentation: https://wcny4qa9krto.feishu.cn/wiki/F4PDw76uDiOG42k76gGc6FaBnod
• English: DataFlow-WebUI Documentation: https://wcny4qa9krto.feishu.cn/wiki/SYELwZhh9ixcNwkNRnhcLGmWnEg

🛠️ 8.5.2 Development Repository

• https://github.com/OpenDCAI/DataFlow-webui

🧪 9. Experimental Results

9.1 DataFlow-Instruct-10k

DataFlow-Instruct-10K is a unified multi-domain instruction dataset generated by the DataFlow framework. It is constructed through several automated data preparation pipelines spanning mathematical reasoning, code, and general text instructions. Each pipeline follows a generate–evaluate–filter–refine workflow to synthesize and curate high-quality instruction–response pairs. The resulting dataset contains approximately 10K samples and provides high-quality supervision for instruction tuning, enabling base models to approach the performance of fully trained instruct models with significantly fewer training examples.

For Detailed Experiments setting, please visit our DataFlow Technical Report.

Model	Math-Avg	Code-Avg	Knowledge-Avg
Qwen2-7B Series
Base	20.1	66.3	76.2
+ Infinity-Instruct-10K	29.0	67.8	76.2
+ Infinity-Instruct-1M	27.9	68.2	76.2
+ DataFlow-Instruct-10K	32.4	66.2	76.1
Qwen2.5-7B Series
Base	37.1	76.5	76.0
+ Infinity-Instruct-10K	22.6	77.6	75.8
+ Infinity-Instruct-1M	33.3	78.0	75.8
+ DataFlow-Instruct-10K	46.7	78.6	76.2

🛠️ 9.2 Other Pipeline Results (Click to expand)

9.2.1 Text Pipeline

9.2.1.1 Pre-training data filter pipeline

From the SlimPajama-627B corpus, we extract a 100B-token subset and apply multiple DataFlow text-pretraining filters. We train a Qwen2.5-0.5B model from scratch for 30B tokens using the Megatron-DeepSpeed framework, the results are as follows:

Methods	ARC-C	ARC-E	MMLU	HellaSwag	WinoGrande	Gaokao-MathQA	Avg
Random-30B	25.26	43.94	27.03	37.02	50.99	27.35	35.26
Qurating-30B	25.00	43.14	27.50	37.03	50.67	26.78	35.02
FineWeb-Edu-30B	26.45	45.41	27.41	38.06	50.43	25.64	35.57
DataFlow-30B	25.51	45.58	27.42	37.58	50.67	27.35	35.69

9.2.1.2 SFT data filter and synthesis pipeline

To study small-scale SFT data quality, we fine-tune the Qwen2.5-7B base model using LLaMA-Factory on WizardLM and Alpaca datasets. For each dataset, we compared a randomly sampled set of 5K instances against a set of 5K instances filtered by DataFlow’s SFT pipeline. Additionally, we synthesize a 15k-size dataset, DataFlow-SFT-15K, using DataFlow’s Condor Generator and Condor Refiner pipeline, followed by DataFlow’s SFT filtering pipeline (excluding the Instagram filter). Benchmarks include comprehensive Math, Code, and Knowledge evaluation suites.

9.2.2 Math Benchmarks

Methods	math	gsm8k	aime24	minerva	olympiad	Avg
Alpaca (random)	54.9	77.2	13.3	14.0	27.0	37.3
Alpaca (filtered)	60.3	80.0	13.3	14.7	30.7	39.8
WizardLM (random)	61.1	84.2	6.7	18.0	29.3	39.9
WizardLM (filtered)	69.7	88.8	10.0	19.9	35.4	44.8
DataFlow-SFT-15K (random)	72.6	89.6	13.3	37.9	32.9	49.3
DataFlow-SFT-15K (filtered)	73.3	90.2	13.3	36.0	35.9	49.7

9.2.3 Code Benchmarks

Methods	HumanEval	MBPP	Avg
Alpaca (random)	71.3	75.9	73.6
Alpaca (filtered)	73.8	75.7	74.8
WizardLM (random)	75.6	82.0	78.8
WizardLM (filtered)	77.4	80.4	78.9
DataFlow-SFT-15K (random)	79.9	75.9	77.9
DataFlow-SFT-15K (filtered)	82.9	74.9	78.9

9.2.4 Knowledge Benchmarks

Methods	MMLU	C-EVAL	Avg
Alpaca (random)	71.8	80.0	75.9
Alpaca (filtered)	71.8	80.0	75.9
WizardLM (random)	71.8	79.2	75.5
WizardLM (filtered)	71.9	79.6	75.8
DataFlow-SFT-15K (random)	72.1	80.0	76.1
DataFlow-SFT-15K (filtered)	72.2	80.4	76.3

9.2.5 Conversation Synthesis Pipeline

We synthesize DataFlow-Chat-15K using DataFlow’s conversation-generation pipeline and fine-tune Qwen2.5-7B-Base on it. Baselines include ShareGPT-15K, UltraChat-15K, and their full (non-truncated) versions. We evaluate on domain-specific tasks (TopDial, Light) and general benchmarks (MMLU, AlpacaEval, Arena-Hard).

9.2.5.1 Conversation Benchmarks

Model	TopDial	Light	Avg
Qwen2.5-7B	7.71	7.79	7.75
+ ShareGPT-15K	7.75	6.72	7.24
+ UltraChat-15K	7.72	6.83	7.28
+ DataFlow-Chat-15K	7.98	8.10	8.04

9.2.5.2 General Benchmarks

Model	MMLU	AlpacaEval	Arena-Hard	Avg
Qwen2.5-7B	71.45	7.05	0.60	26.36
+ ShareGPT-15K	73.09	3.70	1.30	26.03
+ UltraChat-15K	72.97	3.97	0.80	25.91
+ DataFlow-Chat-15K	73.41	10.11	1.10	28.21

9.2.6 Reasoning Pipeline

We adopt the NuminaMath dataset as a high-quality seed dataset. We compare three training sources: (1) a random 10K subset from Open-R1, (2) a random 10K subset from Synthetic-1, and (3) our 10K synthesized DataFlow-Reasoning-10K dataset constructed using DataFlow.

Setting	Model	gsm8k	math	amc23	olympiad	gaokao24_mix	minerva	AIME24@32	AIME25@32	Avg
Baseline	Qwen2.5-32B-Instruct	95.8	73.5	70.0	38.5	42.9	26.5	16.8	11.6	46.95
1 Epoch	+ SYNTHETIC-1-10k	92.9	71.8	52.5	38.4	23.1	24.3	35.6	34.0	46.6
1 Epoch	+ Open-R1-10k	91.5	72.3	65.0	38.4	20.9	24.6	43.0	33.5	48.7
1 Epoch	+ DataFlow-Reasoning-10K	93.9	72.3	72.5	38.7	38.5	26.5	35.9	34.5	51.6
2 Epochs	+ SYNTHETIC-1-10k	94.5	78.4	75.0	45.0	24.2	28.3	48.4	37.9	54.0
2 Epochs	+ Open-R1-10k	93.9	77.2	80.0	44.1	20.9	25.4	51.0	40.7	54.2
2 Epochs	+ DataFlow-Reasoning-10K	94.4	76.6	75.0	45.2	42.9	25.7	45.4	40.0	55.7

9.2.7 Code Pipeline

We randomly sample 20k instances from the Ling-Coder-SFT corpus and process them through the DataFlow Code Pipeline. This yields three curated code instruction datasets of different scales, DataFlow-Code-1K, DataFlow-Code-5K, and DataFlow-Code-10K, each designed to provide high-quality, pipeline-refined supervision signals for code generation tasks. We compare our synthesized datasets against Code-Alpaca-1k and Self-OSS-Instruct-SC2-Exec-Filter-1k.

9.2.7.1 Trained on Qwen2.5-7B-Instruct

Training Data	BigCodeBench	LiveCodeBench (v6)	CruxEval (I)	CruxEval (O)	HumanEval+	Avg
Qwen2.5-7B-Instruct	35.3	23.4	44.8	43.9	72.6	44.0
+ Code Alpaca-1K	33.3	18.7	45.6	46.4	66.5	42.1
+ Self-OSS	31.9	21.4	46.9	45.9	70.1	43.2
+ DataFlow-Code-1K	35.5	25.7	48.0	45.1	72.6	45.4
+ DataFlow-Code-5K	36.2	26.4	48.6	45.0	73.2	45.9
+ DataFlow-Code-10K	36.8	26.0	48.8	45.4	73.8	46.2

9.2.7.2 Trained on Qwen2.5-14B-Instruct

Training Data	BigCodeBench	LiveCodeBench (v6)	CruxEval (I)	CruxEval (O)	HumanEval+	Avg
Qwen2.5-14B-Instruct	37.5	33.4	48.0	48.5	74.4	48.4
+ Code Alpaca-1K	37.0	28.2	50.2	49.6	71.3	47.3
+ Self-OSS	36.9	22.3	52.6	50.1	68.3	46.0
+ DataFlow-Code-1K	41.4	33.7	51.0	50.9	77.3	50.9
+ DataFlow-Code-5K	41.1	33.2	52.5	50.6	76.2	50.7
+ DataFlow-Code-10K	41.9	33.2	52.9	51.0	76.2	51.0

📄 10. Publications

Our team has published the following papers that form core components of the DataFlow system:

Paper Title	DataFlow Component	Venue	Year
DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI	Unified data preparation and workflow automation	ACM SIGMOD / PACMMOD	2027
AgenticRAGTracer: A Clear and Stepwise-Process Benchmark for Agentic RAG	Agentic RAG Data Synthesis	ACL Findings	2026
Text2SQL-Flow: A Robust SQL-Aware Data Augmentation Framework for Text-to-SQL	Text2SQL Data Augmentation	ICDE	2026
Let's Verify Math Questions Step by Step	Math question quality evaluation	KDD	2026
MM-Verify: Enhancing Multimodal Reasoning with Chain-of-Thought Verification	Multimodal reasoning verification framework for data processing and evaluation	ACL	2025
Efficient Pretraining Data Selection for Language Models via Multi-Actor Collaboration	Multi-actor collaborative data selection mechanism for enhanced data filtering and processing	ACL	2025

Contributing Institutions:

🏆 11. Awards & Achievements

We are honored to have received first-place awards in two major international AI competitions, recognizing the excellence and robustness of DataFlow and its reasoning capabilities:

Competition	Track	Award	Organizer	Date
ICML 2025 Challenges on Automated Math Reasoning and Extensions	Track 2:Physics Reasoning with Diagrams and Expressions	🥇First Place Winner	ICML AI for Math Workshop & AWS Codabench	July 18, 2025
2025 Language and Intelligence Challenge (LIC)	Track 2:Beijing Academy of Artificial Intelligence	🥇First Prize	Beijing Academy of Artificial Intelligence (BAAI) & Baidu	August 10, 2025

💐 12. Acknowledgements

We sincerely thank MinerU for their outstanding work, whose powerful PDF/document text extraction capabilities provided essential support for our data loading process. We also thank LLaMA-Factory for offering an efficient and user-friendly framework for large model fine-tuning, which greatly facilitated rapid iteration in our training and experimentation workflows. Our gratitude extends to all contributors in the open-source community—their efforts collectively drive the development of DataFlow. We thank Zhongguancun Academy for their API and GPU support.

🤝 13. Community & Support

Join the DataFlow open-source community to ask questions, share ideas, and collaborate with other developers!

• 📮 GitHub Issues: Report bugs or suggest features

• 🔧 GitHub Pull Requests: Contribute code improvements

• 💬 Join our community groups to connect with us and other contributors!

📜 14. Citation

The DataFlow paper has been accepted for publication in the Proceedings of the ACM on Management of Data (PACMMOD) and presentation at ACM SIGMOD 2027. Until the official ACM citation becomes available, please cite the arXiv version:

@article{liang2025dataflow,
  title={DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI},
  author={Liang, Hao and Ma, Xiaochen and Liu, Zhou and Wong, Zhen Hao and Zhao, Zhengyang and Meng, Zimo and He, Runming and Shen, Chengyu and Cai, Qifeng and Han, Zhaoyang and others},
  journal={arXiv preprint arXiv:2512.16676},
  year={2025}
}