Hybrid Graphs for Table-and-Text based Question Answering using LLMs.
| Authors | Ankush Agarwal et al. |
| Year | 2025 |
| Venue | NAACL 2025 |
| Paper | View on DBLP |
Abstract
Abstract not yet available in this stub. Read the full paper →
Engineering Breakdown
Plain English
This paper addresses the challenge of answering questions that require reasoning over both tabular data and unstructured text simultaneously—a common scenario in enterprise knowledge bases and documentation systems. The authors propose a hybrid graph representation that structures both tables and text as interconnected nodes, allowing large language models to traverse and reason across modalities more effectively. While the abstract isn't available, the title and venue (NAACL 2025) suggest they demonstrate improvements on table-text QA benchmarks by leveraging graph-based reasoning with LLMs rather than traditional retrieval or parsing approaches. This represents a practical advancement for systems that need to answer complex questions where the answer depends on connecting structured data with contextual information from documents.
Core Technical Contribution
The core novelty is the hybrid graph representation that unifies tables and text into a single semantic structure, enabling LLMs to perform joint reasoning across both modalities without treating them as separate retrieval problems. Rather than maintaining separate retrieval paths for tables versus text (the prior standard approach), the authors construct a unified graph where table cells, row/column headers, and text spans become nodes with explicit relational edges, allowing a single LLM reasoning pass to gather evidence from both sources simultaneously. This is fundamentally different from prior work that either extracted table content as text (losing structure) or maintained separate retrievers for tables and documents. The approach leverages modern LLMs' ability to follow graph traversal instructions, using the graph structure as scaffolding for more accurate multi-modal reasoning.
How It Works
The system takes as input a question, a table, and a corpus of associated text documents. It first constructs a hybrid graph where table nodes include individual cells and headers, text nodes include sentence or passage spans, and edges encode relationships like 'cell belongs to row', 'row contains value', and 'text mentions entity'. The question is then passed to an LLM with graph-aware prompting instructions—essentially telling the model to traverse edges in the graph to find relevant evidence. The LLM uses these structural cues to identify which table columns or rows are relevant, which text passages contain supporting context, and how to combine evidence from both modalities to arrive at an answer. The final output is an answer along with a traced reasoning path through the graph showing which nodes were visited, providing interpretability into how the decision was made.
Production Impact
For teams building QA systems over enterprise data, this approach eliminates the need to maintain separate pipelines for table extraction and text retrieval, reducing engineering complexity and potential inconsistencies between modalities. A production system adopting this approach would construct graphs incrementally as documents and tables are indexed, then handle queries by passing the relevant subgraph context to an LLM—trading some upfront graph construction cost for more accurate single-pass reasoning. The interpretability benefit is significant: reasoning traces through the graph help users and compliance teams understand why an answer was given, which is critical in regulated environments like finance or healthcare. Trade-offs include: graph construction overhead at indexing time, larger context windows needed to represent subgraphs (increasing LLM latency and cost), and dependency on the LLM's ability to follow graph traversal instructions, which varies by model and fine-tuning. For systems with <10GB of tabular/textual data, this is likely worthwhile; for massive-scale knowledge bases, selective graph expansion strategies would be needed.
Limitations and When Not to Use This
The approach inherently depends on LLM instruction-following capability, meaning smaller or less capable models may fail to traverse the graph correctly or may be confused by complex graph structures, limiting deployment to frontier models. It assumes that relevant answers can be found by traversing a single connected subgraph—for questions requiring reasoning over distant or weakly-connected information sources, the approach may struggle or require expensive graph expansion. The paper likely doesn't address scalability to millions of rows and documents; graph construction and context window management become prohibitive bottlenecks at massive scale. Additionally, the quality of answers depends heavily on graph edge quality—poorly labeled relationships between table cells and text spans will degrade reasoning accuracy, requiring careful engineering of the graph construction pipeline that the paper may underspecify.
Research Context
This work builds on a decade of table QA research (starting with WikiTableQuestions and HybridQA benchmarks) and extends recent trends in using LLMs for structured reasoning by combining them with explicit graph representations. It bridges table-question-answering and knowledge graph reasoning literatures, applying graph-prompted reasoning (similar to in-context learning over graph paths) to the multimodal setting. The contribution is timely given the 2024-2025 shift toward reasoning-aware LLMs (chain-of-thought, retrieval-augmented generation with structured context), positioning hybrid graphs as a natural scaffolding mechanism for such reasoning. Future work likely extends toward dynamic graph construction (building graphs on-the-fly for queries), graph pruning strategies for scalability, and evaluation on more complex reasoning patterns requiring multi-step graph traversal.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::