Debate to Deceive: Poisoning Large Language Model Fact-Checking with Game-Theoretic Multi-Agent Debate

Researchers propose a knowledge attack framework against large language model fact-checking, Debate-to-Deceive (D2D), which generates semantically coherent, rhetorically effective adversarial text that avoids early confirmation bias through multi-agent debate, in order to manipulate fact-checking judgments. The framework formalizes the adversarial process as an evolutionary game, in which agents dynamically select Logos, Pathos, and Ethos rhetorical moves through evolutionary learning in a compact strategy space, and then a large language model generates concrete text based on the strategy and debate context. On four fact-checking benchmarks, FEVER, HealthVer, CLIMATE-FEVER, and SciFact, D2D significantly outperforms existing attacks such as adversarial suffixes, template injection, and single-agent self-reflection, and bypasses most defenses.