What's Left Unsaid? Detecting and Correcting Misleading Omissions in Multimodal News Previews

Researchers at the National University of Singapore (NUS) have unveiled OMGuard, a novel system designed to detect and correct misleading omissions in multimodal news previews—those image-headline pairs commonly shared on social media. Even when factually accurate, these previews can cause interpretation drift by selectively omitting crucial context, leading readers to form judgments that diverge from the full article. The team first built the MM-Misleading benchmark to simulate how preview-based understanding differs from context-based understanding, revealing pronounced blind spots in open-source large vision-language models (LVLMs) when detecting omission-based misleadingness.

OMGuard operates in two stages: Interpretation-Aware Fine-Tuning (IAFT) for detection, and Rationale-Guided Misleading Content Correction (RGMCC) for rewriting headlines to reduce misleading impressions. Experiments show that OMGuard lifts an 8B parameter model's detection accuracy to match that of a 235B model, while delivering significantly stronger end-to-end correction. Further analysis reveals that misleadingness often stems from local narrative shifts (e.g., missing background) rather than global frame changes, and identifies image-driven cases where text-only correction fails, underscoring the need for visual interventions.