MANTIS: Multi-Axis News Transformations for Instance-Wise Stress Testing and Worst-Case Training in Fake News Detection

Dexin Liu, Fan Yang, Zimu Wang, Xunfei Zhu, Bolin Ren, Bingyan Zhang, Wei Wang, Qi ChenarXiv 2026 · arXiv

TL;DR

MANTIS introduces a new paradigm for robust fake news detection by moving beyond simple "style" checks to a multi-axis adversarial approach. By generating diverse, fact-preserving rewrites of news articles and training models to handle the "worst-case" presentation for every individual instance, the researchers significantly reduce a detector's vulnerability to cosmetic modifications without adding any overhead during actual use.

What Problem Does This Paper Address?

Current automated fake news detectors are surprisingly fragile. They often suffer from shortcut learning, where the model latches onto superficial cues—like an angry tone, tabloid-style phrasing, or sensationalist framing—rather than verifying the underlying factual claims.

This creates a massive security hole: a malicious actor can use a Large Language Model (LLM) to take a piece of fake news and rewrite it in a professional, neutral, "mainstream" style. As shown in Figure 1, these simple cosmetic changes can easily flip a detector's prediction from "Fake" to "Real," even though the deceptive core claim remains identical.

Presentation-level vulnerability: Different styles and tones can trick a detector into misclassifying the same factual claim.

Existing defenses usually focus on a single axis (like just changing the writing style), leaving models blind to other types of manipulation, such as emotional intensification or dissemination framing.

Methods

The authors propose a two-part solution: MANTIS-Bench and the MANTIS Framework.

MANTIS-Bench

They created a benchmark that pairs original news articles with six "transformed views" across three axes:

  1. Writing Style: Mainstream-news vs. Tabloid.
  2. Emotional Expression: Neutral vs. Intensified.
  3. Dissemination Framing: Public-information vs. Engagement-oriented.

These views are generated using LLMs (like GPT-4o-mini) under strict "claim-preservation" constraints to ensure the facts don't change, only the "packaging."

The MANTIS Framework

MANTIS employs an instance-wise worst-case training strategy. Instead of training on all rewrites equally, the framework identifies which specific transformation is most "harmful" (causes the highest loss) for a given news item.

Figure 2: Overview of MANTIS. An LLM rewriter generates six offline views covering writing style, emotional expression, and dissemination framing. MANTIS trains jointly on the original news and each instance’s worst-case view, while inference requires only a single forward pass.

As illustrated in Figure 2, the model is trained on a joint loss: one part focuses on the original article, and the other focuses on the "worst-case" transformed view. Crucially, because these views are generated offline, the final trained detector requires only a single forward pass at inference time—meaning there is zero extra computational cost when deploying the model.

Datasets

The researchers evaluated MANTIS on three major datasets:

  • PolitiFact: Political news (English).
  • GossipCop: Entertainment news (English).
  • Weibo21: A multi-topic Chinese dataset used to test how well the English-trained models generalize across languages and domains.

Key Insights and Ideas

  • Vulnerability is Instance-Specific: The most striking finding is that different news items are weak in different ways. As shown in Figure 4, one article might be easily fooled by a style change, while another is only vulnerable to emotional shifts. A "one-size-fits-all" defense is therefore destined to fail.
Figure 4: Case-specific vulnerability on GossipCop (Llama-3.1-8B-Instruct). Points denote originally cor- rect cases misclassified under one transformation; axes show gold-label probabilities under the harmful and other transformations.
  • The "Union" Metric Matters: The authors introduce the Union Attack Success Rate (UASR). This measures the probability that at least one type of transformation will trick the model. This is a much tougher and more realistic stress test than looking at average accuracy.
  • LLMs are Fragile Detectors: Even powerful zero-shot LLMs (like Llama-3 or Qwen) show significant performance drops when news is rewritten, proving that scale alone doesn't solve the robustness problem.

Key Results

MANTIS consistently outperformed both standard supervised encoders (like RoBERTa) and specialized single-axis defenses (like SheepDog).

  • Robustness Gains: On the GossipCop dataset, MANTIS reduced the UASR by 3.55 percentage points compared to the previous state-of-the-art (SheepDog).
  • Cross-Lingual Transfer: When trained on English data and tested on Chinese news (Weibo21), MANTIS reduced the UASR from 29.42% to 18.80%, showing that learning to be robust to presentation shifts helps the model focus on more universal, fact-based features.
  • Efficiency: MANTIS achieves these gains with a significantly smaller deployment footprint and lower inference costs compared to LLM-based agents that require multiple reasoning steps.

Limitations and Future Work

While MANTIS is a significant step forward, the authors acknowledge several limitations:

  • Finite Transformations: The model only trains on six specific types of rewrites. Real-world attackers might find "gaps" between these axes.
  • Generation Artifacts: While human and LLM-as-a-judge evaluations were high, LLM-generated text can sometimes contain subtle artifacts that might influence training.
  • Multimodality: The current work is text-only. Future iterations could extend this "multi-axis" approach to images and videos, which are increasingly used in sophisticated misinformation campaigns.

Summary updated

Abstract

Automated fake news detection is vital on social media, where fake news is easily rewritten to fool detectors. However, existing approaches typically rely on superficial writing cues or single-style text rewrites, resulting in poor robustness under diverse cosmetic modifications. We construct MANTIS-Bench, a multi-axis presentation-attack benchmark that pairs original news with rewritten views across writing style, emotional expression, and dissemination framing. We further propose MANTIS, an instance-adaptive adversarial framework enabling dynamic worst-case training tailored to each news article with zero inference-time overhead. This reframes fake news defense from heuristic format smoothing into targeted, fact-invariant credibility assessment. Experiments on MANTIS-Bench show that MANTIS reduces the Union Attack Success Rate (UASR) by 3.55 percentage points relative to SheepDog on GossipCop and consistently outperforms prior single-axis defenses on Macro-F1 while maintaining competitive clean accuracy, demonstrating the effectiveness of multi-axis and instance-adaptive worst-case training.
BibTeX
@misc{liu2026mantis,
  author = {Dexin Liu and Fan Yang and Zimu Wang and Xunfei Zhu and Bolin Ren and Bingyan Zhang and Wei Wang and Qi Chen},
  title = {{MANTIS: Multi-Axis News Transformations for Instance-Wise Stress Testing and Worst-Case Training in Fake News Detection}},
  year = {2026},
  eprint = {2610.35475},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.35475},
}
Figure