MERIT: Modular Framework for Multimodal Misinformation Detection with Web-Grounded Reasoning

Technical Summary: MERIT (Modular Framework for Multimodal Misinformation Detection)

1. Problem Statement

The rapid spread of multimodal misinformation—where false narratives are constructed through the coordinated manipulation of text and images—outpaces human fact-checking capabilities. Existing detection methods face three primary challenges:

• Generalization Gap: Supervised models require massive, domain-specific datasets and struggle to adapt to new types of manipulation.
• Knowledge Limitations: Large Vision-Language Models (LVLMs) often lack real-time factual knowledge, leading to "hallucinated" explanations or inconsistent judgments.
• Architectural Limitations: Most current systems use end-to-end prompting, which fails to decompose the complex, multi-faceted nature of multimodal verification (e.g., distinguishing between a fake image, a false claim, and a mismatched context).

2. Methodology

The authors propose MERIT (Multimodal Evidence-based Retrieval for Image-Text verification), an inference-time modular framework. Instead of a single prompt, MERIT decomposes the verification process into four sequential, specialized modules:

1. Visual Verification Module: Analyzes the image for technical artifacts (e.g., warped anatomy, inconsistent lighting) and contextual anomalies (e.g., physics violations) to detect AI generation or manipulation.
2. Relevancy Assessment Module: Evaluates the semantic alignment between the image and the text headline, categorizing the relationship as true, partial, or false to detect out-of-context misinformation.
3. Claim Verification Module (RAG-based): Employs a multi-stage Retrieval-Augmented Generation (RAG) process. It generates investigative questions across three reasoning chains, retrieves web evidence via DuckDuckGo, and synthesizes citation-linked answers to verify the textual claim.
4. Final Judgment Module: Integrates signals from the previous three modules using structured decision rules. To classify content as "Not Misinformation," the system requires all three signals to pass (factually supported headline, aligned image, and non-AI-generated imagery).

The framework is model-pluggable, meaning it can utilize any instruction-following LVLM (the authors used GPT-4o-mini) as its reasoning engine.

3. Key Contributions

• Modular Architecture: A novel decomposition of multimodal verification that separates visual forensics, cross-modal alignment, and factual retrieval.
• Evidence-Based Reasoning: The system produces rationales linked to web citations, enhancing transparency and human auditability.
• Controlled Evaluation: The authors conducted a "same-model" evaluation (comparing MERIT against a re-implemented MMD-Agent using the same GPT-4o-mini model) to prove that performance gains come from the architectural design rather than simply using a more powerful model.

4. Results

Evaluated on the MMFakeBench dataset, MERIT demonstrated significant improvements:

• Superior Performance: Achieved an 81.65% F1 score, outperforming the GPT-4V MMD-Agent baseline (74.0% F1).
• Enhanced Recall: MERIT achieved a 6.14 point higher misinformation recall than the MMD-Agent baseline, meaning it is significantly better at catching actual misinformation.
• Module Specialization (Ablation Studies):
  • Removing the Visual Module caused a 36-point drop in detecting visual distortions.
  • Removing the Claim Module caused a major drop in detecting textual distortions.
  • Removing the Relevancy Module primarily degraded performance in cross-modal mismatch detection.
• Generalization: Test set evaluations (5,000 samples) showed results nearly identical to validation results, confirming the framework's robustness.

5. Significance

MERIT represents a shift from "black-box" end-to-end prompting toward structured, tool-augmented reasoning for misinformation detection. Its significance lies in its ability to provide interpretable and verifiable judgments. By breaking down the problem, the framework allows for easier debugging and updates (e.g., upgrading the search engine or the visual detector independently). While the authors note that the current false-positive rate (due to legitimate news using generic stock photos) and reliance on APIs limit real-time deployment, the framework provides a scalable blueprint for human-in-the-loop fact-checking systems.