A new pipeline lets a frozen large language model answer questions that span multiple database tables — without touching the model's weights.
Researchers introduced GRAB, a three-stage system that converts relational data into a heterogeneous graph, runs message-passing to encode the structure, then compresses the result into a small set of query-conditioned "latent tokens" fed to the LLM alongside ordinary text. The LLM itself is never updated; only the graph encoder and the latent bridge — totaling 91 million parameters — are trained. That is a fraction of the cost of fine-tuning or retrieval-augmented approaches that touch the base model.
The gains show up most sharply in multi-table settings, where a model must reason across joins and foreign keys rather than a single flat sheet. That is exactly where standard text-flattening tricks break down, because the relational structure gets lost in serialization. GRAB preserves it through the graph encoding stage, giving the LLM something closer to a schema-aware summary than a wall of comma-separated values.
The 91M-parameter bridge is small enough to retrain for new domains without spinning up a full model fine-tune — which is either an elegant design choice or a very convenient way to describe a system that still depends on a large frozen model you do not control.