Research map / Large language models / Retrieval and knowledge

HeGTa: Leveraging Heterogeneous Graph-enhanced Large Language Models for Few-shot Complex Table Understanding

AAAI 2025 · Large language models › Retrieval and knowledge · approach: Knowledge graph reasoning with LLMs · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HeGTa, an LLM (Vicuna-7B) augmented with a heterogeneous-graph transformer encoder whose outputs are soft prompts, trained via self-supervised instruction tuning and few-shot task fine-tuning for complex table understanding.

Problem

Table understanding under few-shot conditions with complex table structures (merged cells, hierarchical headers), where existing LLM, graph, and BERT-style methods either lose topological information or need large labeled data.

Approach

Convert tables into heterogeneous graphs with heuristic edge rules, encode them with an HGT, and inject the node vectors as soft prompts into a LoRA-tuned LLM, pre-trained with three multi-granularity self-supervised tasks (row classification, cell matching, context generation) before few-shot task tuning.

Key results

• HeGTa reaches 66.39 on HiTab CTC and 59.58 on IM-TQA CTC, outperforming the best baseline (TabPrompt 63.56 and TableLlama 53.40 respectively) in the 8-shot setting.
• On TQA, HeGTa scores 45.12 on WTQ (Flat) and 45.96 on WTQ (Raw), versus 44.38 and 42.97 for COTable.
• Removing the CoT annotations drops WTQ (Flat) accuracy by 2.31 points, and removing the HG encoder (homogeneous graph) reduces performance across most tasks.

Measured effects reported in the paper

Surprising findings

• Removing the TRC objective slightly improves WCC TTC (+1.02) and HiTab-based TQA (+1.19 on IM-TQA), indicating that the row classification task does not help every task.

Limitations

• Evaluated only on 8-shot settings with small training sets, and performance on tables with more diverse layouts is not explored.
• Weaker on textual table QA reasoning than program-aided methods; the authors note integrating program-aided paradigms as future work.
• Heuristic graph construction rules and header detection are hand-designed.

Experimental setup

Datasets: WCC IM-TQA HiTab WTQ TabFact TURL

Benchmarks: HiTab IM-TQA WCC WTQ (Flat) WTQ (Raw) TabFact

Metrics: Macro-F1 accuracy

Baselines: ForTap GetPt TabularNet TabPrompt Binder Aug-Codex COTable TableCoT TableLlama

Builds on: LLaVA Time-LLM GraphGPT Heterogeneous Graph Transformer Auto-CoT

Closest papers by meaning

• NeedleInATable: Exploring Long-Context Capability of Large Language Models towards Long-Structured Tables (NeurIPS 2025 · similarity 0.88)
• TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.87)
• TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy (NeurIPS 2024 · similarity 0.86)
• Table as a Modality for Large Language Models (NeurIPS 2025 · similarity 0.86)
• DiGraphHal-Bench: Evaluating Multimodal Large Language Models on Complex Directed Graphs (CVPR 2026 · similarity 0.86)
• Decoupling Skeleton and Flesh: Efficient Multimodal Table Reasoning with Disentangled Alignment and Structure-aware Guidance (ICML 2026 · similarity 0.85)