Latent Bridges for Multi-Table Question Answering

A new pipeline called GRAB tackles multi-table question answering by converting relational data into a heterogeneous graph and feeding a compact structural signal to a frozen large language model, according to research posted on arXiv [1]. The constructor-encoder-bridge pipeline lifts rows, column classes, and value groups into typed nodes, with foreign-key metadata inducing shared column classes across tables [3]. A graph encoder then processes this structure via message passing to capture row–column–value dependencies and cross-table connectivity [3]. The encoded graph is projected into a fixed-length sequence of soft tokens through a query-conditioned latent bridge, and those tokens are injected as a prefix before the text prompt that the LLM receives [3]. The LLM itself remains frozen throughout; only the graph encoder and the bridge—totaling 91M parameters—are updated during training [1]. The approach departs from methods that rely solely on flattened table text. Instead of a single static summary, the bridge produces a question-relevant structural representation [3]. The projected latent vectors are mapped to the LLM’s embedding dimension by a learned projector before being prepended to the prompt [4]. The researchers report the largest performance gains in demanding multi-table settings [2]. Multi-table question answering has drawn increasing attention in the research community. A separate framework, Graph-Table-RAG, reorganizes table corpora into a heterogeneous hypergraph and uses coarse-to-fine retrieval before applying graph-aware prompting for downstream reasoning [9]. Another recent study constructs synthetic contrastive reasoning traces for multi-table Q&A, generating positive traces that derive the gold answer and negative traces that contain plausible but incorrect reasoning; fine-tuning open-weight LLMs with Contrastive Preference Optimization on these pairs yielded absolute average improvements of 9.7%–16.3% over standard supervised fine-tuning across three model families [10]. The GRAB pipeline offers a complementary direction by keeping the LLM frozen and offloading structural encoding to a lightweight trainable module [1]. The latent-bridge design echoes the broader concept of latent semantic analysis, where a compressed representation captures relational structure, though LSA historically relied on singular value decomposition over term-document matrices rather than learned graph encoders [6]. The work was posted on arXiv on 27 June 2026 [1].