Abstract: Large Language Model (LLM) equipped with strong parametric knowledge and reasoning capabilities has shown great potential in fake news detection, yet its internal decision-making robustness remains largely unexplored. Observations show that LLM displays stable reasoning trajectories similar to those of human fact-checkers, which is termed cognitive path preference. Path Preference Evaluation Framework (PPEF) based on "LLM as a Judge" paradigm is first proposed to formalize and quantify this phenomenon. PPEF first extracts explanatory reasoning rationales of detectors, maps them to preference feature ontology, and further identifies path preferences shared by multiple LLMs. To further examine whether these preferences can be exploited as vulnerabilities by malicious propagators, Multi-View Rewrite (MVR) strategy is designed that selectively weakens cues aligned with LLM preferences while keeping original news semantics intact. Experiments across multiple public datasets and different detection models show that LLM exhibits clear and consistent dependence on a small set of cues during fake news detection. Compared to five non-preference-oriented rewriting methods including Style-Based Rewriting (SBG) and Open-ended Generation (OEG), MVR targeting preference features leads to more severe performance degradation of detectors. Specifically, on Gossipcop dataset, adversarial samples generated by MVR decrease detection rates of basic CoT-driven LLM (Vanilla LLM) by 1.8 to 46.7 percentage points, demonstrating significantly superior interference effectiveness over baseline methods like SBG and OEG. These results reveal a practically exploitable LLM vulnerability and offer insights for developing preference-aware and path-diversified detection frameworks.

Key words: Multi-View Rewrite (, MVR)