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

TL;DR

This work formalizes the problem of misleading omissions in multimodal news previews and introduces the MM-Misleading benchmark (6,000 instances) to study detection and correction. It proposes OMGuard, a two-stage framework combining interpretation-aware fine-tuning and rationale-guided mitigation to detect preview-context misalignment and rewrite headlines to reduce misleading impressions, while keeping images as anchors when necessary. Experiments show an 8B model with interpretation-aware training can match or outperform a 235B LVLM in detection and achieve strong end-to-end correction, though image-driven omissions reveal limits of text-only edits and motivate multimodal interventions. The findings emphasize that most misleadingness arises from local narrative gaps rather than global frame changes, and that visual cues or image editing strategies can be crucial for robust governance of misleading previews in practice.

Abstract

Even when factually correct, social-media news previews (image-headline pairs) can induce interpretation drift: by selectively omitting crucial context, they lead readers to form judgments that diverge from what the full article conveys. This covert harm is harder to detect than explicit misinformation yet remains underexplored. To address this gap, we develop a multi-stage pipeline that disentangles and simulates preview-based versus context-based understanding, enabling construction of the MM-Misleading benchmark. Using this benchmark, we systematically evaluate open-source LVLMs and uncover pronounced blind spots to omission-based misleadingness detection. We further propose OMGuard, which integrates (1) Interpretation-Aware Fine-Tuning, which used to improve multimodal misleadingness detection and (2) Rationale-Guided Misleading Content Correction, which uses explicit rationales to guide headline rewriting and reduce misleading impressions. Experiments show that OMGuard lifts an 8B model's detection accuracy to match a 235B LVLM and delivers markedly stronger end-to-end correction. Further analysis reveals that misleadingness typically stems from local narrative shifts (e.g., missing background) rather than global frame changes, and identifies image-driven scenarios where text-only correction fails, highlighting the necessity of visual interventions.

Figures (18)

• Figure 1: Illustration of misleading omissions in multimodal news previews. Social media users typically see only a news preview (image–headline pair), while the full context becomes available only after clicking through. When key information is omitted or selectively presented, the preview can induce misinterpretations that diverge from those supported by the full article.
• Figure 2: Overview of OMGuard: the upper section shows the multi-stage annotation pipeline described in § \ref{['sec:data_annotation']}; the lower section presents OMGuard, where the model is first fine-tuned with interpretation-aware supervision using misleadingness rationales, and then applied to rationale-guided correction of misleading previews.
• Figure 3: Quantifying error propagation via oracle substitution of $U_p/U_c$; $\Delta=\text{Acc}(\text{oracle }U)-\text{Acc}(\text{self }U)$.
• Figure 4: Frame-shift analysis comparing stylistic preservation and frame alignment in misleading headline correction.
• Figure 5: Statistic of different misleading types.