Latent Bridges for Multi-Table Question Answering

2026-06-27

AI Summary

This work addresses the challenge of effectively integrating structured multi-table relational data with large language models (LLMs) for complex question answering. The authors propose GRAB, a novel framework that constructs a heterogeneous graph to represent relational data and employs graph neural networks for encoding. A lightweight, query-conditioned implicit bridging module then injects compact structural representations into a frozen LLM without requiring fine-tuning. Notably, GRAB introduces the first implicit bridging mechanism, achieving strong synergy between structured data and language models while training only a 91M-parameter auxiliary subnetwork. Experimental results demonstrate that GRAB significantly outperforms existing methods on challenging multi-table QA benchmarks, with particularly pronounced gains in high-difficulty scenarios.

Abstract

We introduce GRAB, a constructor-encoder-bridge pipeline for table question answering. Our method lifts relational data into an heterogeneous graph, encodes it via message passing, and transfers the signals to an LLM through a small set of query-conditioned latent tokens. This provides the LLM with a compact, task-relevant structural representation together with the flattened text. Crucially, the LLM remains strictly frozen to preserve its general reasoning capabilities; we train only the lightweight graph encoder and latent bridge (91M parameters), allowing the entire pipeline to be trained efficiently. Our pipeline significantly improves performance on relational Question Answering, with the largest gains in demanding multi-table settings, offering an efficient, principled way to connect relational deep learning with LLMs.

Problem

• multi-table question answering
• relational data
• large language models
• structured representation

Innovation

• latent bridge
• heterogeneous graph
• frozen LLM
• multi-table QA
• message passing