CMIE: Combining MLLM Insights with External Evidence for Explainable Out-of-Context Misinformation Detection

Fanxiao Li, Jiaying Wu, Canyuan He, Wei Zhou

Introduction

In the contemporary digital landscape, the most insidious forms of misinformation often leverage authentic media in deceptive ways. Out-of-Context (OOC) misinformation—where a genuine, unaltered image is paired with a fabricated or misleading caption—presents a significant challenge for automated detection systems. While Multimodal Large Language Models (MLLMs) like GPT-4o possess sophisticated reasoning capabilities, they frequently succumb to "deceptive correlations," where the superficial plausibility of an image-text pair masks an underlying falsehood.

This paper introduces CMIE (Combining MLLM Insights with External Evidence), a framework designed to enhance the detection of OOC misinformation by compelling models to analyze the semantic relationship between media components before evaluating external evidence. By integrating a novel "Association Scoring" mechanism, CMIE achieves state-of-the-art accuracy while providing transparent, human-readable explanations for its verdicts.

The Challenge of Out-of-Context Detection

Detecting OOC misinformation is inherently more complex than identifying manipulated media (e.g., "deepfakes") because the visual content itself is authentic. The deception resides entirely in the relationship between the image and the text. Current AI approaches suffer from two primary limitations:

  1. Superficial Reasoning: MLLMs tend to be "conservative" in their internal knowledge. If an image and text share a common thematic background—such as a courtroom or a political rally—the model often defaults to a "Real" label, failing to scrutinize the specific claims being made.
  2. The Noise Problem in Retrieval: While Retrieval-Augmented Generation (RAG) allows models to access external facts, web search results are often cluttered with irrelevant or tangentially related information. Without a mechanism to distinguish critical evidence from background noise, the inclusion of external data can actually degrade model performance.

The CMIE Framework

CMIE addresses these gaps through a multi-stage pipeline that prioritizes semantic alignment and evidence relevance:

1. Evidence Collection

The system utilizes the Google Vision API to extract key entities from the image and employs web crawlers to identify page titles and contexts where the image has previously appeared. This provides a factual foundation for the subsequent analysis.

2. Coexistence Relationship Generation (CRG)

Rather than immediately attempting a verdict, the MLLM is tasked with identifying the underlying semantic links that might justify why the specific image and text appear together. The model generates a "coexistence score" and a rationale, forcing it to articulate the presumed logic of the pairing before verifying it against external facts.

3. Association Scoring (AS)

Using the output from the CRG stage as a filter, the model evaluates each piece of retrieved external evidence. It assigns a relevance score to each snippet, ensuring that the final judgment is informed by "smoking gun" evidence rather than irrelevant news snippets. This step is crucial for mitigating the "noise problem" inherent in RAG.

4. Final Prediction and Explanation

The model integrates the original image, the text, the filtered evidence, and the coexistence analysis to produce a final classification (Real or Fake). Crucially, it generates a detailed, human-readable explanation that justifies the decision.

Experimental Results

Performance on NewsCLIPpings

The researchers evaluated CMIE on the NewsCLIPpings benchmark, the largest dataset for OOC detection, specifically focusing on the "Merged/Balanced" subset which contains "hard negatives"—cases where the image and fake caption are semantically related.

  • Accuracy: CMIE achieved a state-of-the-art accuracy of 91%, significantly outperforming the previous leading model, SNIFFER (88%), and standard CLIP-based approaches (65%).
  • Precision: The framework achieved a precision of 0.93 on fake samples, indicating a remarkably low false-alarm rate.
  • Ablation Study: The inclusion of the Association Scoring (AS) mechanism alone improved accuracy from 87% to 90%, validating the hypothesis that filtering evidence is as critical as the retrieval itself.

Explainability

Beyond binary classification, CMIE provides qualitative value. Human evaluators rated CMIE’s explanations at 3.9/5 for accuracy and comprehensiveness, surpassing other state-of-the-art models like LEMMA (3.5/5). This transparency is vital for building user trust in automated fact-checking systems.

Limitations and Future Directions

While CMIE represents a significant advancement, the authors note several areas for improvement. Like all LLM-based systems, CMIE is susceptible to occasional hallucinations or misinterpretations of complex visual cues. Furthermore, the multi-stage pipeline incurs higher computational costs than single-pass models. Future work will explore the use of "intelligent agents" to cross-check explanations and optimize the framework for real-time deployment, further narrowing the gap between AI reasoning and human logic.

Summary updated

BibTeX
@inproceedings{li2025cmie,
  author = {Fanxiao Li and Jiaying Wu and Canyuan He and Wei Zhou},
  title = {{CMIE: Combining MLLM Insights with External Evidence for Explainable Out-of-Context Misinformation Detection}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.findings-acl.487},
}
Figure