GPE: Robust Evidence Aggregation for Fact-Checking under Controllable GEO-Style Poisoning

The study proposes GPE, a multi-domain fact-checking benchmark and evaluation framework for evaluating the ability of fact-checking methods to resist GEO-style poisoning under controllable evidence sources and controllable poisoning ratios. The authors point out that large language models obtain information via search tools, and manipulated retrieved documents constitute a new attack surface; generative engine optimization increases the probability that specific content is retrieved, cited, and adopted. Experiments across multiple verification methods and poisoning attacks show that GPE can expose robustness degradation and efficiency trade-offs that are invisible only in clean evaluations.