E2LVLM:Evidence-Enhanced Large Vision-Language Model for Multimodal Out-of-Context Misinformation Detection

TL;DR

E2LVLM tackles multimodal out-of-context misinformation by aligning external textual evidence with LVLM inputs through a two-step process of textual evidence reranking and rewriting. It follows with a one-stage instruction-tuning regime that couples judgment and explanation generation, built atop an open-source LVLM, to achieve superior detection accuracy and compelling rationales. The approach is validated on NewsCLIPpings, where E2LVLM surpasses previous state-of-the-art methods and demonstrates strong ablations, explanation quality, and robustness on real-world data such as VERITE. This work offers a practical, explainable pathway for rapid, evidence-aware debunking of multimodal misinformation in real-world deployment.

Abstract

Recent studies in Large Vision-Language Models (LVLMs) have demonstrated impressive advancements in multimodal Out-of-Context (OOC) misinformation detection, discerning whether an authentic image is wrongly used in a claim. Despite their success, the textual evidence of authentic images retrieved from the inverse search is directly transmitted to LVLMs, leading to inaccurate or false information in the decision-making phase. To this end, we present E2LVLM, a novel evidence-enhanced large vision-language model by adapting textual evidence in two levels. First, motivated by the fact that textual evidence provided by external tools struggles to align with LVLMs inputs, we devise a reranking and rewriting strategy for generating coherent and contextually attuned content, thereby driving the aligned and effective behavior of LVLMs pertinent to authentic images. Second, to address the scarcity of news domain datasets with both judgment and explanation, we generate a novel OOC multimodal instruction-following dataset by prompting LVLMs with informative content to acquire plausible explanations. Further, we develop a multimodal instruction-tuning strategy with convincing explanations for beyond detection. This scheme contributes to E2LVLM for multimodal OOC misinformation detection and explanation. A multitude of experiments demonstrate that E2LVLM achieves superior performance than state-of-the-art methods, and also provides compelling rationales for judgments.

Figures (6)

• Figure 1: Subfigures (a) and (b) show the representation distributions between images and their retrieved/rewritten evidence. In addition, subfigure (c) provides a quantization comparison of E2LVLM over the OOC on NewsCLIPpings luo2021newsclippings.
• Figure 2: An overview of E2LVLM - the evidence-enhanced large vision-language model. Given an authentic image input together with a text claim, Google APIs are used to retrieve external evidence about the image-claim pair in an inverse search manner abdelnabi2022open. The image and its retrieved textual evidence are input to a Large Vision-Language model (Qwen2-VL) for evidence reranking. The top-1 textual evidence is further rewritten by the LVLM model. Apart from textual evidence, the retrieved visual evidence about the claim is reranked by cosine similarity, which achieves the top-1 visual evidence. Such content is input to E2LVLM together with the task-specific prompt for desired behaviors. Given this context, E2LVLM can provide its judgment and explanation for the authenticity of the image-claim pair.
• Figure 3: Prompts and their examples in E2LVLM. (a) Reranking prompt $\mathcal{P}_\mathrm{rerank}$ is to select one most relevant textual evidence related to the authentic image. (b) Rewriting prompt $\mathcal{P}_\mathrm{rewrite}$ is to achieve the coherent and contextually attuned content for alignment. (c) Explanation prompt $\mathcal{P}_\mathrm{Expla.}$ is to generate the compelling rationale to make up support for its assessment. (d) Tuning prompt $\mathcal{P}_\mathrm{OOC}$ is to extend the general-purpose LVLM to the task of multimodal out-of-context misinformation detection.
• Figure 4: Comparison of E2LVLM with various reranking ways.
• Figure 5: Visualization of various data distributions.