TL;DR
StructGPT introduces an Iterative Reading-then-Reasoning (IRR) framework that enables LLMs to query tables, knowledge graphs, and databases through specialized reading interfaces without any fine-tuning.
Summary
LLMs cannot directly query structured data sources, and single-shot context stuffing loses structure and exceeds context limits. StructGPT addresses this with an Iterative Reading-then-Reasoning (IRR) loop: specialized reading interfaces extract relevant rows, triples, or schema elements from tables, KGs, or databases, which are passed as context to the LLM for a reasoning step; the loop repeats until an answer is produced. Evaluated on WikiTableQuestions, WebQSP, and BIRD, it rivals fine-tuned baselines across all three structured QA settings without any model fine-tuning.
Key contributions
- Proposes the IRR (Iterative Reading-then-Reasoning) loop as a unified framework for LLM reasoning over heterogeneous structured data.
- Introduces modular reading interfaces specialized for tables, knowledge graphs, and SQL databases under a single architecture.
- Demonstrates that iterative retrieval over structured sources rivals fine-tuned models without any gradient updates to the LLM.
- Shows the framework generalizes across WikiTableQuestions, WebQSP (KG), and BIRD (database) benchmarks.
When to cite
- When enabling an LLM to reason over structured data sources (tables, KGs, databases) without fine-tuning.
- When arguing that iterative structured retrieval outperforms single-shot context stuffing for complex QA.
- When designing modular reading interfaces for heterogeneous data formats under a unified LLM reasoning framework.
- When comparing zero-shot LLM approaches to fine-tuned baselines on structured question answering.