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

TL;DR

The paper investigates how multimodal large language models can detect out-of-context misinformation and identifies limitations in relying solely on internal knowledge or loosely connected external evidence. It introduces CMIE, a framework combining a Coexistence Relationship Generation (CRG) strategy with an Association Scoring (AS) mechanism to uncover deeper image-text relationships and selectively leverage external evidence. Through extensive experiments on NewsCLIPpings, CMIE outperforms strong baselines and provides human-readable explanations, achieving around 0.91 accuracy and robust performance across prompts and model variants. The work advances explainable, evidence-aware OOC misinformation detection and highlights the importance of structured reasoning over surface-level evidence matching.

Abstract

Multimodal large language models (MLLMs) have demonstrated impressive capabilities in visual reasoning and text generation. While previous studies have explored the application of MLLM for detecting out-of-context (OOC) misinformation, our empirical analysis reveals two persisting challenges of this paradigm. Evaluating the representative GPT-4o model on direct reasoning and evidence augmented reasoning, results indicate that MLLM struggle to capture the deeper relationships-specifically, cases in which the image and text are not directly connected but are associated through underlying semantic links. Moreover, noise in the evidence further impairs detection accuracy. To address these challenges, we propose CMIE, a novel OOC misinformation detection framework that incorporates a Coexistence Relationship Generation (CRG) strategy and an Association Scoring (AS) mechanism. CMIE identifies the underlying coexistence relationships between images and text, and selectively utilizes relevant evidence to enhance misinformation detection. Experimental results demonstrate that our approach outperforms existing methods.

Figures (12)

• Figure 1: Comparison of different paradigms for MLLM-based OOC misinformation detection.
• Figure 2: Example of direct reasoning in OOC misinformation detection using MLLM's internal knowledge.
• Figure 3: Examples of augmented reasoning by integrating external evidence.
• Figure 4: Overall architecture of the proposed CMIE, which effectively models the relationship between external evidence, generating predictive labels and interpretable explanations.
• Figure 5: Performance comparison of GPT-4o and Gemini-Pro under evidence-augmented reasoning and CMIE.