DIVER: Dynamic Iterative Visual Evidence Reasoning for Multimodal Fake News Detection

Weilin Zhou, Zonghao Ying, Chunlei Meng, Jiahui Liu, Hengyang Zhou, Quanchen Zou, Deyue Zhang, Dongdong Yang, Xiangzheng ZhangarXiv 2026 · arXiv

TL;DR

DIVER is a dynamic framework for multimodal fake news detection that mimics human cognitive processing by separating fast textual filtering from deep visual analysis. By selectively triggering fine-grained visual tools (like OCR and dense captioning) only when cross-modal contradictions arise, it improves detection accuracy by an average of 2.72% while significantly reducing inference latency and token consumption.

What Problem Does This Paper Address?

Multimodal fake news detection (MFND) aims to identify misinformation by analyzing both text and accompanying images. However, current state-of-the-art approaches suffer from two main flaws: computational redundancy and visual hallucination.

Most existing models are static—they apply the same heavy visual processing to obvious truths and subtle fabrications alike, wasting massive computational resources. Furthermore, indiscriminately fusing global visual embeddings often injects background noise or causes Large Language Models (LLMs) to hallucinate false correlations when subtle visual details contradict the text narrative.

Figure 1: Two primary limitations of existing multi- modal methods: (a) noise induced factual ambiguity, and (b) visual hallucination due to weak grounding.

Methods

Inspired by dual-process cognitive theory, the authors propose DIVER (Dynamic Iterative Visual Evidence Reasoning), which splits verification into a fast screening phase (System 1) and a deliberate reasoning phase (System 2).

The pipeline consists of four progressive stages:

  1. Linguistic Investigation: A fine-tuned LLM ($LLM_{analyst}$) decomposes raw text into discrete, atomic factual claims, separating objective facts from stylistic cues (like emotional manipulation).
  2. Intra-modal Consistency Reflection: A verification module checks the faithfulness of these extracted claims against the original text using global and local semantic metrics. If the LLM hallucinated details, a self-correction loop triggers it to try again.
  3. Inter-modal Alignment Verification: A lightweight gate uses CLIP to score the semantic alignment between the verified text claims and the image. If alignment is high, the sample is deemed safe, and the pipeline terminates early (System 1).
  4. Evidence-driven Visual Forensics: If alignment is low, DIVER activates specialized visual tools (OCR, dense captioning, object tagging) based on the semantic category of the claim (System 2). This discrete visual evidence is iteratively fed back to refine the textual claims.

Finally, an Uncertainty-Aware Dynamic Fusion (UDF) module uses a masked attention mechanism to aggregate features, ensuring that skipped visual modalities contribute exactly zero influence rather than introducing zero-padded noise.

Figure 2: Overview of the DIVER framework featuring two explicit feedback loops. The pipeline initiates with linguistic investigation to verify intra-modal consistency. Subsequently, a cross-modal alignment gate dynamically regulates reasoning depth: samples exhibiting high semantic consensus are resolved immediately, while disjoint cases trigger evidence-driven visual forensics, selectively activating visual analysis capabilities (e.g., OCR) to aggregate evidence for claim refinement.

Datasets

The framework was evaluated on three widely used multimodal fake news benchmarks:

  • Weibo: A Chinese social media dataset containing 7,532 training and 1,996 test articles.
  • Weibo21: A newer iteration of the Weibo dataset containing 9,127 total articles across varying topics.
  • GossipCop: An English entertainment news dataset containing 10,010 training and 2,830 testing instances.

No new datasets were introduced in this work.

Key Insights and Ideas

  • Cognitive Economy: Not all news requires deep visual forensics. By filtering out roughly 70% of straightforward samples early via text-image alignment, DIVER saves massive computational overhead.
  • Discrete Evidence vs. Global Embeddings: Passing raw global visual embeddings into an LLM often leads to hallucinations. Extracting discrete, text-based visual evidence (like OCR text from a jersey or specific object tags) bridges the semantic gap far more effectively.
  • The Danger of "Over-thinking": The authors found that setting the self-correction loop iteration limit ($\tau$) too high ($\tau > 2$) actually degraded performance. Excessive iteration causes the model to hallucinate non-existent conflicts, proving that a tight balance between reasoning depth and semantic stability is crucial.

Key Results

  • State-of-the-Art Accuracy: DIVER achieved accuracies of 94.8% on Weibo, 96.2% on Weibo21, and 91.6% on GossipCop, outperforming top baselines (like KEN and MIMoE-FND) by an average of 2.72%.
  • Faster Inference: It reduced average inference latency to 4.12 seconds, outperforming LLM-based baselines like LIFE (7.09s) and INSIDE (4.97s).
  • Token Optimization: DIVER reduced average token consumption to 1.5k per sample, a 6.25% improvement over the most efficient baseline, thanks to its dynamic gating mechanism.

Limitations and Future Work

The authors identify three key areas for future exploration:

  1. Domain Generalization: Enhancing the framework's ability to generalize across more diverse and shifting social media platforms.
  2. Adversarial Robustness: Evaluating scalability in complex, real-world scenarios and testing robustness against unforeseen, highly sophisticated adversarial attacks.
  3. Edge Deployment: Investigating lightweight model distillation techniques to adapt the framework for resource-constrained edge devices.

Summary updated

Abstract

Multimodal fake news detection is crucial for mitigating adversarial misinformation. Existing methods, relying on static fusion or LLMs, face computational redundancy and hallucination risks due to weak visual foundations. To address this, we propose DIVER (Dynamic Iterative Visual Evidence Reasoning), a framework grounded in a progressive, evidence-driven reasoning paradigm. DIVER first establishes a strong text-based baseline through language analysis, leveraging intra-modal consistency to filter unreliable or hallucinated claims. Only when textual evidence is insufficient does the framework introduce visual information, where inter-modal alignment verification adaptively determines whether deeper visual inspection is necessary. For samples exhibiting significant cross-modal semantic discrepancies, DIVER selectively invokes fine-grained visual tools (e.g., OCR and dense captioning) to extract task-relevant evidence, which is iteratively aggregated via uncertainty-aware fusion to refine multimodal reasoning. Experiments on Weibo, Weibo21, and GossipCop demonstrate that DIVER outperforms state-of-the-art baselines by an average of 2.72\%, while optimizing inference efficiency with a reduced latency of 4.12 s.
BibTeX
@misc{zhou2026diver,
  author = {Weilin Zhou and Zonghao Ying and Chunlei Meng and Jiahui Liu and Hengyang Zhou and Quanchen Zou and Deyue Zhang and Dongdong Yang and Xiangzheng Zhang},
  title = {{DIVER: Dynamic Iterative Visual Evidence Reasoning for Multimodal Fake News Detection}},
  year = {2026},
  eprint = {2601.07178},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.07178},
}
Figure