SINT-Flow: Schema Integration using Large Language Model Workflows

2026-07-27 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

SINT-Flow is a novel schema integration framework designed to automate the derivation of a global, unified schema from diverse input tables. Comprising five LLM-based operators, SINT-Flow can process denormalized source tables containing attributes for multiple entity types, decomposing them into distinct entity-specific relations. The framework was evaluated using GPT-5.2 and the open-weight Qwen-3.6-27B model on SINT-Bench, a new benchmark featuring 10 schema integration tasks across 93 relational tables. SINT-Flow achieved F1 scores of at least 96% for entity-type detection, 85% for attribute detection, and 83% for schema mapping. An ablation study further confirmed the effectiveness of its self-consistency strategy and the inclusion of a review loop within the schema matching operator. This work was published on 2026-07-27.

Key takeaway

For Machine Learning Engineers building data pipelines, SINT-Flow demonstrates a robust approach to automated schema integration. If your projects involve consolidating diverse, denormalized datasets, consider implementing LLM-based workflows with decomposition and review loops. This method, achieving high F1 scores with models like GPT-5.2, can significantly streamline data preparation, reducing manual effort and improving data quality for downstream ML tasks.

Key insights

SINT-Flow employs LLM workflows to automate end-to-end schema integration, effectively handling denormalized source tables.

Principles

• LLM operators can automate complex schema integration.
• Decompose denormalized tables for better integration.
• Self-consistency and review loops boost LLM accuracy.

Method

SINT-Flow employs five LLM-based operators in configurable workflows to perform fully automated schema integration, including decomposing denormalized source tables into entity-specific relations.

In practice

• Apply LLM workflows for schema integration.
• Evaluate LLM performance on denormalized data.
• Integrate review loops for accuracy.

Topics

• Schema Integration
• Large Language Models
• LLM Workflows
• Data Denormalization
• SINT-Bench
• Entity Detection

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• ChatGPT, Claude, Gemini :How Large Language Models Are Changing Data Science — Large Language Models are evolving data science by automating routine tasks, enabling data scientists to focus on higher-value, complex problem-solving.
• An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models — LLMs can automate multi-variable control design and tuning through structured code generation and validation.
• Database Normalization via Dual-LLM Self-Refinement — Miffie automates database normalization using a dual-LLM self-refinement loop with specialized zero-shot prompts for accuracy and cost-efficiency.
• Code-on-Graph: Iterative Programmatic Reasoning via Large Language Models on Knowledge Graphs — Code-on-Graph integrates LLMs with KGs using programmatic reasoning, representing schemas as Python classes for flexible, scalable knowledge utilization.
• DataFlow-Harness: A Grounded Code-Agent Platform for Constructing Editable LLM Data Pipelines — DataFlow-Harness bridges the NL2Pipeline gap by enabling LLM agents to construct persistent, editable data pipelines through platform-grounded, incremental mutations.
• Balancing Usefulness and Naturalness: An LLM-based Curation Pipeline for Code Review Comments — Curating LLM training data for code review comments by balancing quality and natural diversity significantly improves model performance.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.