Paper summary: DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI

Problem

LLM data preparation is mostly ad-hoc scripts and loosely specified workflows, hard to reproduce and with little support for model-in-the-loop generation.

Method

DataFlow provides (Figure 1):

• system-level abstractions for modular, reusable, composable data transformations
• a PyTorch-style API for building debuggable, optimisable pipelines
• nearly 200 operators and six general pipelines: text, maths reasoning, code, Text-to-SQL, agentic RAG and large-scale knowledge extraction
• DataFlow-Agent, which translates natural-language specs into executable pipelines through operator synthesis, planning and iterative verification

Key results

From the abstract, across six use cases:

• up to +3% execution accuracy in Text-to-SQL over SynSQL
• +7% average on code benchmarks
• 1–3 point gains on MATH, GSM8K and AIME
• a unified 10K-sample dataset lets base models surpass counterparts trained on 1M Infinity-Instruct samples

Limitations

There’s no dedicated limitations section in the extracted text. Gains are measured on the authors’ six use cases.