Inconsistent Reasoning Attacks to Identify Weaknesses in Automatic Scientific Claim Verification Tools

This chapter explores the vulnerabilities of Scientific Claim Verification (SCV) tools, which are crucial for assessing the validity of scientific claims. The text introduces inconsistent reasoning attacks, a broader category of adversarial manipulations that include specific-to-general, general-to-specific, negation, union, and paraphrase attacks. These attacks systematically expose flaws in SCV decision-making by manipulating claim logic rather than just altering surface-level text. The chapter also presents an evolutionary algorithm that leverages large language models to generate effective adversarial attacks, targeting logical inconsistencies in SCV tools. Evaluation against MultiVerS, one of the most advanced SCV models, reveals persistent vulnerabilities, while a retrieval-augmented generation (RAG) system with an Attack-Reflection mechanism shows promise in mitigating these issues. The chapter concludes by emphasizing the importance of strengthening SCV systems against such attacks to ensure their reliability and suggests future research directions for integrating self-reflection and adversarial robustness to improve SCV effectiveness.