MAGIC: A Multi-Hop and Graph-Based Benchmark for Inter-Context Conflicts in Retrieval-Augmented Generation

Jungyeon Lee, Lee Kangmin, Taeuk KimEMNLP 2025

TL;DR

This paper introduces MAGIC, a novel benchmark for evaluating how well Large Language Models (LLMs) can detect and pinpoint knowledge conflicts within and between different pieces of information. They found that even advanced LLMs struggle significantly with these conflicts, especially when multiple reasoning steps are involved, highlighting a critical gap in current AI capabilities for reliable information integration.

What Problem Does This Paper Address?

Retrieval-Augmented Generation (RAG) systems, which enhance LLMs by providing external knowledge, often face a challenge: the retrieved information might contradict the LLM's internal knowledge or even contradict itself. This "knowledge conflict" can lead to incorrect or unreliable outputs. While existing benchmarks exist, they are often limited to simple question-answering scenarios, rely on basic text manipulation (like entity substitution), and don't adequately capture the complexity of real-world conflicts, especially those requiring multi-hop reasoning. The authors aim to address this by creating a more comprehensive and challenging benchmark.

Methods

The core of their contribution is a framework for generating a new dataset called MAGIC. This framework leverages knowledge graphs (KGs) as a structured source of information. The process involves:

  1. Subgraph Extraction: Distilling relevant parts of a large KG (Wikidata5M) into smaller, manageable knowledge chunks.
  2. Knowledge Conflict Generation: Intentionally introducing conflicts within these subgraphs. This is done by perturbing nodes and edges in the KG. They use LLMs with a few-shot prompting strategy to generate these conflicts, aiming for subtle and indirect contradictions that require multi-hop reasoning.
  3. KG-to-Text Conversion: Converting the original and perturbed KGs into natural language text passages using KG-to-text generation algorithms. This ensures the conflicts are presented in a format that LLMs can process.

A key aspect is a two-stage human-in-the-loop process to ensure high-quality data, filtering out trivial or incoherent generated conflicts.

Datasets

The primary dataset introduced is MAGIC (A Multi-Hop And Graph-Based Benchmark for Inter-Context Conflicts). It's constructed using knowledge graphs and features a balanced distribution of conflict types, including single-hop and multi-hop, as well as single and multiple conflicts (N-conflict).

For comparison, they also evaluate on:

  • ECON (Jiayang et al., 2024): A dataset with 168 instances focusing on evidence conflicts.
  • WikiContradict (Hou et al., 2024): A benchmark of 103 samples derived from Wikipedia, categorizing conflicts as explicit or implicit.

Key Insights and Ideas

  • Multi-hop reasoning is a major hurdle: LLMs, even powerful ones, significantly struggle with detecting conflicts that require chaining multiple pieces of information or relations. This is a critical limitation for real-world applications where information is rarely simple.
  • Localization is harder than detection: Models are better at identifying that a conflict exists than pinpointing where it is within the text. This is crucial for debugging and understanding AI outputs.
  • KG structure aids interpretability: Using KGs allows for a more structured and interpretable way to generate and represent conflicts, offering insights into the underlying relationships that are being contradicted.
  • Dataset complexity matters: MAGIC's balanced distribution of conflict types and its emphasis on multi-hop scenarios make it a more challenging benchmark than existing ones, pushing the boundaries of LLM evaluation.

Key Results

  • Overall Performance: LLMs tested on MAGIC consistently showed lower Identification (ID) and Localization (LOC) scores compared to ECON and WikiContradict, with average ID scores decreasing by up to 10% and LOC scores by up to 17%.
  • Top Performer: GPT-4o-mini achieved the highest accuracy on MAGIC, but still struggled with multi-hop reasoning.
  • Conflict Type Impact: Single-hop conflicts were much easier for LLMs than multi-hop conflicts. The number of conflicts (N) aided identification but hindered localization.
  • Domain Analysis: Models performed best on "Class/Concept" domains and varied widely on "Organization" domains, suggesting differences in handling hierarchical relations.
  • Human Baseline: Human annotators achieved significantly higher ID (92.5%) and LOC (83.3%) scores than current LLMs, underscoring the difficulty of the task.

Limitations and Future Work

  • KG Source: MAGIC is currently built on Wikidata5M. Incorporating additional KGs or domain-specific sources could improve robustness.
  • Localization Evaluation: Current localization evaluation relies on human judgment; developing automated approaches is a future direction.
  • Parametric Knowledge: The study touched upon how LLMs' internal (parametric) knowledge affects conflict detection, suggesting a need for more systematic investigation into this interaction.
  • Optimized Methods: The authors aim to develop optimized methods to help LLMs overcome their current limitations in handling complex knowledge conflicts.

Summary updated

BibTeX
@inproceedings{lee2025magic,
  author = {Jungyeon Lee and Lee Kangmin and Taeuk Kim},
  title = {{MAGIC: A Multi-Hop and Graph-Based Benchmark for Inter-Context Conflicts in Retrieval-Augmented Generation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure