arXiv cs.CL· Simone Varriale, Tamara Cucumides, Floris Geerts, Paolo Papotti·· 13 小时前AI 评分31
GRAB：用潜在桥接让冻结 LLM 处理多表问答
Latent Bridges for Multi-Table Question Answering
AI 导读
GRAB 提出一种 constructor-encoder-bridge 流水线，将关系数据提升为异构图、经消息传递编码后，用少量查询条件化的 latent token 把结构信号传给严格冻结的 LLM，仅训练 91M 参数的轻量图编码器与潜在桥接。该方法在关系型问答上显著提升性能，多表场景增益最大。
正文 · 原文
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.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Databases (cs.DB) |
| Cite as: | arXiv:2606.28916 [cs.CL] |
| (or arXiv:2606.28916v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2606.28916 arXiv-issued DOI via DataCite |
Submission history
From: Simone Varriale [view email]
[v1]
Sat, 27 Jun 2026 13:48:20 UTC (2,228 KB)
[v2]
Wed, 21 Oct 2026 15:06:47 UTC (410 KB)
来源：arXiv cs.CL · arxiv.org