💬 NLP

iQUEST: An Iterative Question-Guided Framework for Knowledge Base Question Answering

The paper introduces iQUEST, an iterative question-guided framework that combines sub-question decomposition with Graph Neural Network-enhanced look-ahead reasoning to improve multi-hop Knowledge Base Question Answering performance across various benchmarks and Large Language Models.

Original authors: Shuai Wang, Yinan Yu

Published 2026-02-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shuai Wang, Yinan Yu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a very tricky riddle, but you don't have all the answers in your head. You need to look them up in a giant, messy library (the Knowledge Graph).

The problem with current "smart" computers (Large Language Models, or LLMs) is that when they try to solve these riddles, they often get lost. They might wander down the wrong aisle, pick up a book that looks similar but isn't right, or give up too early because the answer isn't on the very first shelf they check.

This paper introduces iQUEST, a new way to help these computers solve complex riddles. Think of iQUEST not as a single genius, but as a team of detectives working together with a map.

Here is how it works, broken down into simple parts:

1. The "Break It Down" Strategy (Iterative Question Guidance)

Imagine you are asked: "What is the official flower of the area hit by Tropical Storm Fabio?"

If you ask a computer this directly, it might get confused. It sees "Storm," "Flower," and "Area," and tries to jump straight to the answer. It might guess "Mexico" because storms happen there, but miss that the specific area is Hawaii.

iQUEST changes the game: Instead of asking the big question all at once, it teaches the computer to ask itself smaller, easier questions first, like a detective following a trail of clues.

Step 1: "Where did Tropical Storm Fabio hit?" (Answer: Hawaii).
Step 2: "What is the official flower of Hawaii?" (Answer: Yellow Hibiscus).

By breaking the big problem into tiny steps, the computer stays focused and doesn't get lost in the noise. It builds a clear path from the question to the answer.

2. The "Look Ahead" Glasses (GNN & 2-Hop Neighbors)

This is the paper's second superpower.

Imagine you are looking for a specific person in a crowd.

Old Way: You only look at the people standing immediately next to the person you are asking about. If the person you need isn't right there, you give up.
The iQUEST Way: The computer puts on "Look Ahead Glasses." It doesn't just look at the people standing next to the target; it looks at the people standing next to those people (two steps away).

Why does this matter?
Sometimes the answer isn't directly connected to the first clue.

Example: You are looking for the composer of a movie.
Clue 1: The movie title.
Clue 2 (1-step away): The director. (Maybe the director isn't the composer).
Clue 3 (2-steps away): The composer is connected to the director, but not directly to the movie title in the database.

Old methods might stop at the director and say, "Nope, wrong person." iQUEST looks two steps ahead, sees the composer, and says, "Aha! That's the connection we were missing!" This prevents the computer from throwing away the right answer too early.

3. The Teamwork (Two Models)

The system uses two different "brains" to do the job:

The Guide (IQG-LLM): This is the detective who asks the small questions and decides where to look next. It doesn't need to know the final answer; it just needs to know how to navigate the library.
The Answerer (AE-LLM): This is the expert who reads the books the Guide found and writes the final answer.

The Result

The authors tested this system on four different "test exams" (datasets) using various computer brains. The results were impressive:

It solved more riddles correctly than previous methods.
It worked well even with smaller, cheaper computer models.
It was especially good at solving "multi-hop" riddles (riddles that require connecting several dots).

In a Nutshell

iQUEST is like giving a smart computer a step-by-step checklist and a telescopic lens.

The checklist stops it from getting overwhelmed by big questions.
The lens lets it see connections that are two steps away, so it doesn't miss the answer just because it wasn't standing right next to the first clue.

This makes the computer much more reliable when it needs to find facts in a huge, complex world of data.

1. Problem Definition

The paper addresses the challenges of Knowledge Base Question Answering (KBQA), specifically focusing on complex, multi-hop reasoning tasks. While Large Language Models (LLMs) excel in general NLP, they often suffer from hallucinations and factual inconsistencies in knowledge-intensive domains. Integrating Knowledge Graphs (KGs) offers a solution, but reasoning over KGs presents two critical bottlenecks:

Loss of Reasoning Coherence: Models struggle to maintain a logical trajectory when navigating multiple hops, often getting distracted by ambiguous or tangential entities (noise).
Premature Discarding of Critical Paths: Existing methods often rely on 1-hop neighbors (immediate connections) to select the next step. This can lead to the early elimination of valid paths where the crucial connection lies in 2-hop neighbors (e.g., an entity with low immediate relevance but a highly relevant 2-hop connection).

2. Methodology: iQUEST Framework

The authors propose iQUEST, an iterative, question-guided framework that decomposes complex queries into simpler sub-questions while leveraging a Graph Neural Network (GNN) for forward-looking exploration. The framework consists of three core modules:

A. Iterative Question Guidance (IQG-LLM)

Instead of decomposing the entire question once at the start, iQUEST dynamically generates a new sub-question at each reasoning step based on the current context ( $C$ ) and the original question ( $Q$ ).

Process: The LLM (IQG-LLM) analyzes the current state and generates a specific sub-question ( $Q_{sub}$ ) that is fully answerable from the current context.
Goal: This maintains a focused reasoning trajectory, preventing the model from losing direction in complex multi-hop scenarios.
Loop: The process iterates: Generate Sub-question $\rightarrow$ Retrieve Evidence $\rightarrow$ Answer Sub-question $\rightarrow$ Update Context $\rightarrow$ Check Sufficiency.

B. Two-Hop Entity Exploration (GNN-Based)

To address the issue of premature path discarding, the framework moves beyond simple 1-hop retrieval.

Neighbor Retrieval: The system retrieves 1-hop neighbors of the current topic entity via SPARQL queries.
2-Hop Aggregation: It then retrieves the 2-hop neighbors of those 1-hop nodes.
GNN Scoring: A Graph Neural Network (specifically a GraphSAGE-like architecture) aggregates semantic information from the 2-hop neighbors to update the representation of the 1-hop neighbors.
- Mechanism: The representation of a 1-hop node ( $h_{1h}$ ) is updated by concatenating it with an aggregated representation of its 2-hop neighbors ( $h_{2h}$ ).
- Classification: A Multi-Layer Perceptron (MLP) classifies these updated nodes as "relevant" or "irrelevant" based on the current sub-question.
Outcome: This "look-ahead" mechanism ensures that entities with strong 2-hop connections are not discarded due to weak 1-hop relevance scores.

C. Answer Extraction (AE-LLM)

Once the most relevant entities are selected, a second LLM (AE-LLM) synthesizes the answer.

It takes the sub-question and the retrieved context (entities and relations) as input.
It generates an intermediate answer, which is added to the context history.
If the accumulated context is sufficient to answer the original question, the AE-LLM synthesizes a final response; otherwise, the loop continues.

3. Key Contributions

Iterative Question-Guided Reasoning: Unlike prior work that performs a single decomposition, iQUEST iteratively generates sub-questions at every step. This creates a structured, self-correcting reasoning path that adapts to the evolving context.
GNN-Enhanced Forward-Looking Search: The integration of a GNN to aggregate 2-hop neighbor information allows the model to anticipate critical connections that 1-hop methods miss, significantly improving the robustness of multi-hop reasoning.
Comprehensive Evaluation: The framework was tested across four benchmark datasets (CWQ, WebQSP, WebQuestions, GrailQA) and four different LLMs (GPT-4o, DeepSeek-R1, LLaMA 70B, LLaMA 3B), demonstrating consistent improvements without requiring fine-tuning of the LLMs themselves.

4. Experimental Results

The paper reports state-of-the-art (SOTA) or near-SOTA performance across multiple benchmarks:

Performance: iQUEST achieved the highest scores on WebQSP (88.93%) and WebQuestions (81.23%), and ranked second on CWQ (73.85%) and GrailQA (73.52%).
Comparison: It significantly outperformed baselines like ToG and Interactive-KBQA, even when those baselines utilized powerful models via API.
Ablation Studies:
- KG Augmentation: Removing the KG caused a substantial performance drop, confirming the necessity of external structured knowledge.
- GNN Impact: The GNN-based 2-hop retrieval consistently improved performance across all model configurations, proving its value in filtering noise and finding hidden paths.
- LLM Roles: The study revealed that the Answer Extraction (AE-LLM) benefits more from strong internal knowledge (e.g., GPT-4o), while the Question Guidance (IQG-LLM) benefits from strong reasoning capabilities. However, even a smaller model (LLaMA 3B) could effectively guide the process if paired with a strong AE-LLM.

5. Significance and Limitations

Significance:

Efficiency & Adaptability: iQUEST offers a powerful alternative to fine-tuning LLMs for specific domains. By training a lightweight GNN for entity selection and using LLMs for reasoning/guidance, it achieves high adaptability with lower computational costs than full model fine-tuning.
Robustness: The framework effectively solves the "hallucination" problem in KBQA by grounding reasoning in verifiable KG paths while using LLMs to navigate the complexity of multi-hop queries.

Limitations:

Computational Overhead: The dual-LLM architecture (one for guidance, one for answering) introduces latency and cost, which may hinder real-time applications.
Fixed Hop Depth: The GNN is currently limited to 2-hop neighbors. While sufficient for many cases, it may struggle with extremely deep reasoning chains in specific domain KGs where 3+ hops are required. Future work suggests dynamic path search strategies.

In conclusion, iQUEST represents a significant advancement in KBQA by combining the structured reasoning of iterative sub-questions with the semantic depth of GNN-based graph exploration, effectively bridging the gap between LLM capabilities and structured knowledge retrieval.