A Logic-Grounded Language Model Paradigm For Relational Data

Objectif

"Relational databases power finance, healthcare and science, yet they are not natively understood by today’s Large Language Models (LLMs). LLMs lack relational understanding: they are trained on text and, to consume tables in a prompt, they must convert them into a sequence of words, destroying the connections and rules - such as joins, keys and constraints - that carry logic in relational data. Scaling monolithic models is not the answer: it dilutes relational structure, demands prohibitive energy, and forces organizations to expose sensitive data to external clouds.

The GENESIS project confronts this challenge by establishing a formal theory of Relational Grounding, a new paradigm that defines how relational logic can be intrinsically represented in language models. This theory underpins our approach of Grounded Self-Supervision. Instead of retrofitting LLMs for relational data, our framework uses an organization's database - schema, constraints, and content - as a blueprint to automatically derive training examples grounded in the database's logic. These examples guide the creation of a specialized Small Language Model, trained once and operating on live relational data, designed to ""think in tables"" and achieve superior, verifiable performance. Rather than scaling a single, monolithic LLM, our approach enables the automated, on-demand construction of multiple frugal, expert models. By embedding relational logic into the model architecture, instead of learning it by brute force, we drastically reduce parameter count and energy costs.

GENESIS will redefine how organizations interact with their data by replacing traditional query tools with conversational, auditable systems. Its principles will seed the creation of specialized, sustainable language models beyond relational data. This project will position Europe at the forefront of sovereign, climate-friendly, and fact-grounded generative AI, empowering experts to reason over their data with regulatory compliance."