Beyond Retrieval: Improving Evidence Quality for LLM-based Multimodal Fact-Checking

TL;DR

This work tackles the challenge of multimodal disinformation verification with LLM-based systems constrained by training data cutoffs and noisy external evidence. It introduces Aletheia, an end-to-end framework that separates retrieval-oriented multimodal claim interpretation from a structured evidence quality evaluation pipeline. The evidence pipeline extracts eight factual dimensions, filters by credibility, and scores evidence with a reframed score q_i = \alpha \cdot r_i + (1-\alpha) \cdot m_i, enhancing coverage and reducing noise to improve verification. Empirical results on Mocheg, MR2, and the open-world MMDV demonstrate strong accuracy gains (e.g., 88.3% on MR2 and 90.2% open-world) and superior efficiency relative to baselines like DEFAME, underscoring the critical role of evidence quality in reliable multimodal fact-checking. The work points to practical impact in automated fact-checking and sets a path for future extension to audio/video modalities and more robust evidence sourcing.

Abstract

The increasing multimodal disinformation, where deceptive claims are reinforced through coordinated text and visual content, poses significant challenges to automated fact-checking. Recent efforts leverage Large Language Models (LLMs) for this task, capitalizing on their strong reasoning and multimodal understanding capabilities. Emerging retrieval-augmented frameworks further equip LLMs with access to open-domain external information, enabling evidence-based verification beyond their internal knowledge. Despite their promising gains, our empirical study reveals notable shortcomings in the external search coverage and evidence quality evaluation. To mitigate those limitations, we propose Aletheia, an end-to-end framework for automated multimodal fact-checking. It introduces a novel evidence retrieval strategy that improves evidence coverage and filters useless information from open-domain sources, enabling the extraction of high-quality evidence for verification. Extensive experiments demonstrate that Aletheia achieves an accuracy of 88.3% on two public multimodal disinformation datasets and 90.2% on newly emerging claims. Compared with existing evidence retrieval strategies, our approach improves verification accuracy by up to 30.8%, highlighting the critical role of evidence quality in LLM-based disinformation verification.

Figures (6)

• Figure 1: Comparisons of different strategies (including our Aletheia) for multimodal fact-checking.
• Figure 2: Overview of Aletheia, improving evidence quality for multimodal fact-checking by (1) retrieval-oriented multimodal claim interpretation and (2) structured evidence quality evaluation.
• Figure 3: LLM strategies for verifying disinformation.
• Figure 4: Representative examples in our empirical study. Analysis denotes the manual analysis of the failure reason. The incorrect contents of the answers generated by the LLMs are highlighted in red.
• Figure 5: An illustrative example of how Aletheia automatically verifies the real-world multimodal misinformation.