Relational database synthesis stops at numbers and categories: REMIX also generates text columns and cross-table rules, cutting constraint violations by about half
Synopsis
Relational database synthesis has largely covered numerical and categorical columns, leaving text columns and cross-table semantic rules handled separately. REMIX builds the database as a typed heterogeneous graph, tunes a small language model through graph-to-prompt mapping so it generates mixed-type records under relational context, and has a large language model discover and validate semantic dependencies that are compiled into tested repair programs. Across seven real relational databases, REMIX lowers the average constraint violation rate by at least 49.11% relative to baselines on four text-free databases and by at least 47.69% on three text-rich ones.
Interpretation
Relational database synthesis now brings text columns and cross-table semantic rules into a single generation pass instead of only fitting numerical and categorical distributions. Earlier statistical, GAN, autoregressive, diffusion and language-model methods natively supported mostly numerical and categorical attributes, with text either generated separately or excluded. On four text-free databases REMIX has the lowest constraint violation rate on all four, an average relative reduction of at least 49.11%; on three text-rich databases it reduces the rate by at least 47.69% on average relative to GReaT-Cond and GReaT-Joint.
Generation runs in two steps: a small language model writes records under relational context, then programs repair values that violate semantic dependencies. The small model is conditioned not on serialized parent records or joined wide tables but on soft prompts, continuous vectors aggregated from a heterogeneous graph. Record nodes are encoded by column type with column name and value embedded together, a two-layer heterogeneous graph neural network aggregates connected records, and a projector yields one soft prompt token; the large model proposes at most 30 candidate dependencies and validates them over up to 5 rounds on fresh samples before compiling them into repair programs.
Acceptance of semantic dependencies carries a statistical guarantee rather than case-by-case human judgment. Asking a large language model to enumerate dependencies directly yields an incomplete or inaccurate set, and source noise makes valid rules occasionally violated. Each candidate is tested with a one-sided Clopper–Pearson upper confidence bound, with a violation tolerance of 0.1 and a total error budget of 0.1, so accepted dependencies have violation probability at most the tolerance with at least the stated confidence.
Perspective
Evaluation covers seven real relational databases: four text-free ones (Walmart, Biodeg., F1 without text, HM) and three text-rich ones (Chicago, F1 with text, ML-small); synthesis keeps the source schema, table sizes and primary/foreign-key assignments fixed. The authors list scaling to larger relational databases and more downstream machine-learning tasks as future work.
On the text-rich databases, ML-small TSTR mean absolute error is 2.30% higher than GReaT-Joint, while REMIX leads on the other metrics and databases. Larger decoder size does not consistently improve fidelity, and Qwen3-0.6B gives the lowest C2ST-Agg on five of seven databases; temperature and soft-prompt token count produce only moderate differences.