Diagnosing and Addressing Pitfalls in KG-RAG Datasets: Toward More Reliable Benchmarking

Liangliang Zhang, Zhuorui Jiang, Hongliang Chi, Haoyang Chen, Mohammed Elkoumy, Fali Wang, Qiong Wu, Zhengyi Zhou, Shirui Pan, Suhang Wang, Yao Ma

Introduction

Knowledge Graph Retrieval-Augmented Generation (KG-RAG) has emerged as a powerful technique to mitigate hallucinations in Large Language Models (LLMs) by grounding responses in structured, factual data. To evaluate these systems, researchers have traditionally relied on standard Knowledge Graph Question Answering (KGQA) benchmarks. However, as models advance, the reliability of these legacy benchmarks has come into question. This research identifies critical flaws in existing datasets and proposes a new framework for generating high-fidelity, multi-hop benchmarks.

The Benchmark Quality Crisis

A comprehensive manual audit of 16 widely used KGQA datasets, including WebQSP and ComplexWebQuestions (CWQ), reveals a startling lack of data integrity. The research team found that the average factual correctness across these benchmarks is only 57%. Issues range from outdated ground truths and ambiguous questions to fundamentally unanswerable queries. Furthermore, the traditional "Exact Match" (EM) evaluation metric often penalizes correct answers that differ lexically from the ground truth (e.g., "AUD" vs. "Australian dollar").

Table 1: Factual correctness audit of 16 popular KGQA datasets, showing significant error rates in legacy benchmarks.

These findings suggest that high performance on legacy benchmarks may provide a false sense of progress, as the datasets themselves are often too noisy or simple to challenge modern LLM-based systems.

The KGQAGen Framework

To address these pitfalls, the authors developed KGQAGen, an automated, LLM-in-the-loop pipeline designed to produce verifiable, multi-hop QA pairs. The framework operates in four distinct stages:

  1. Seed Initialization: The process begins with a seed entity from Wikipedia’s Vital Articles, used to extract an initial local subgraph from Wikidata.
  2. LLM-Guided Expansion: To ensure complexity, an LLM evaluates the subgraph. If it lacks the depth required for multi-hop reasoning, the LLM identifies specific entities for further expansion, avoiding the noise of blind graph traversal.
  3. Question Generation: The LLM generates a natural language question, its corresponding answer, and the minimal supporting subgraph.
  4. Symbolic Verification: Crucially, the system generates a SPARQL query for each question. This query is executed against the live Wikidata knowledge graph. A QA pair is only retained if the SPARQL results match the LLM-generated answer, effectively eliminating hallucinations.
Figure 1: The KGQAGen pipeline, illustrating the iterative process of subgraph expansion, question generation, and symbolic verification.

KGQAGen-10k and Semantic Evaluation

Using this pipeline, the authors constructed KGQAGen-10k, a dataset comprising 10,787 high-quality, multi-hop QA pairs. Manual verification of a sample subset confirmed a factual accuracy of 96.3%, a significant improvement over legacy benchmarks.

To solve the evaluation rigidity problem, the authors introduced LLM-Assisted Semantic Match (LASM). This metric uses a lightweight LLM to determine if a predicted answer is semantically equivalent to the ground truth, providing a more accurate reflection of model performance than string-based matching.

Key Results and the Retrieval Gap

Experiments on KGQAGen-10k reveal that state-of-the-art KG-RAG models struggle significantly more than previously thought. Models that achieved 75-92% accuracy on older datasets saw their performance drop to the 52-54% range on the new benchmark.

A critical insight from this study is the "retrieval gap." While GPT-4o achieves 54.2% accuracy on KGQAGen-10k in a standard RAG setting, its performance jumps to 84.9% when provided with the "gold" supporting subgraph. This indicates that while LLMs possess strong reasoning capabilities over structured data, the retrieval components of current KG-RAG pipelines are the primary bottleneck.

Conclusion

The transition toward more reliable KG-RAG systems requires benchmarks that are both factually accurate and structurally complex. By identifying the systemic failures in legacy datasets and providing the KGQAGen framework, this work sets a new standard for benchmarking. Future efforts will focus on refining the SPARQL verification process and improving the robustness of entity labeling to further enhance dataset quality.

Summary updated

BibTeX
@inproceedings{zhang2025diagnosing,
  author = {Liangliang Zhang and Zhuorui Jiang and Hongliang Chi and Haoyang Chen and Mohammed Elkoumy and Fali Wang and Qiong Wu and Zhengyi Zhou and Shirui Pan and Suhang Wang and Yao Ma},
  title = {{Diagnosing and Addressing Pitfalls in KG-RAG Datasets: Toward More Reliable Benchmarking}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=Vd5JXiX073},
}
Figure