gtfintechlab/KG-MuLQA-D

KG-MuLQA: A Framework for KG-based Multi-Level QA Extraction and Long-Context LLM Evaluation

KG‑MuLQA is a framework that (1) extracts QA pairs at multiple complexity levels (2) along three key dimensions -- multi-hop retrieval, set operations, and answer plurality, (3) by leveraging knowledge-graph-based document representations.

Overview of KG-MuLQA. Credit agreements are annotated to identify entities and their relationships, forming a knowledge graph representation. This graph is then used to systematically extract multi-level QA pairs, which serve as the basis for benchmarking long-context LLMs.

KG‑MuLQA-D Dataset

We produce KG‑MuLQA‑D, a dataset of 20,139 QA pairs derived from 170 SEC credit agreements (2013–2022) and categorized by five complexity levels. Each QA pair is tagged with a composite complexity level (L = #hops + #set‑ops + plurality), split into Easy, Medium, and Hard.

This table illustrates the question templates used to construct KG-MuLQA-D, structured along three dimensions: plurality (P), number of hops (H), and set operations (#SO). It includes example templates, corresponding knowledge graph query paths, and logical operations involved. These dimensions are used to compute the overall complexity level for each QA pair. The full list of templates can be found in the paper.

LLM Benchmarking & Evaluation

We evaluate 16 proprietary and open-weight LLMs on KG-MuLQA-D benchmark. As question complexity increases, the LLM's ability to retrieve and generate correct responses degrades markedly. We categorize observed LLM failures into four major types, each of which presents recurring challenges as question complexity increases: Misinterpretation of Semantics, Implicit Information Gaps, Set Operation Failures, and Long-Context Retrieval Errors. See the paper for detailed analysis.

This table presents the performance of 16 LLMs, evaluated across Easy, Medium, and Hard question categories. The metrics include the F1 Score and the LLM-as-a-Judge rating, capturing both token-level accuracy and semantic correctness. The results reveal a consistent decline in performance as question complexity increases, with notable model-specific strengths and weaknesses. * denotes the models evaluated on a smaller subset due to cost constraints (see the paper for extended evaluation).

Dataset Release

To facilitate reproducibility and future research, we release the KG‑MuLQA‑D dataset under a CC-BY-NC-ND 4.0 license. The dataset is divided into development and test sets as follows:

Stats	Dev	Test	Total
# Documents	40	130	170
# Questions per Doc (Min)	1	1	1
# Questions per Doc (Avg)	14.75	23.49	21.44
# Questions per Doc (Max)	83	428	428
# Easy Questions	1,499	5,051	6,550
# Medium Questions	2,680	10,203	12,883
# Hard Questions	239	467	706
Total Questions	4,418	15,721	20,139

• Development Set (~25%): 40 documents and 4,418 QA pairs are publicly released to support model development and validation.
• Test Set (~75%): 130 documents and 15,721 QA pairs are not released to prevent data contamination and ensure fair evaluation (questions are released for the leaderboard).

Citation

If you use KG‑MuLQA in your work, please cite:

@misc{tatarinov2025kgqagenknowledgegraphbasedframeworksystematic,
      title={KG-QAGen: A Knowledge-Graph-Based Framework for Systematic Question Generation and Long-Context LLM Evaluation}, 
      author={Nikita Tatarinov and Vidhyakshaya Kannan and Haricharana Srinivasa and Arnav Raj and Harpreet Singh Anand and Varun Singh and Aditya Luthra and Ravij Lade and Agam Shah and Sudheer Chava},
      year={2025},
      eprint={2505.12495},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.12495}, 
}

For questions or issues, please reach out to:

• Nikita Tatarinov: ntatarinov3@gatech.edu
• Agam Shah: ashah482@gatech.edu