EMNLP 2026, Conference on Empirical Methods in Natural Language Processing, 24-29 October 2026, Budapest, Hungary
Despite their widespread adoption, the robustness of fact-checking models to adversarial perturbations remains underexplored. Existing approaches are model-specific and require gradient access or dataset-dependent optimization, with implications for generalization, efficiency, and semantic validity. We introduce FactFlip, a framework for analyzing robustness in claim verification models via universal adversarial triggers. FactFlip identifies perturbative trigger words through a lightweight, model-only procedure that ranks candidates by the alignment of their non-contextualized embeddings with class-specific decision directions, without labeled data or gradient access. Trigger discovery is decoupled from claim perturbation via an LLM-based perturb-and-verify pipeline that preserves semantic validity. Experiments show that FactFlip discriminates between strong and weak triggers, composes them constructively, and rivals fully supervised gradient-based baselines. We further show that FactFlip doubles as a diagnostic lens, surfacing socially salient model biases without any task-specific supervision.
Type:
Conference
City:
Budapest
Date:
2026-10-24
Department:
Data Science
Eurecom Ref:
9034
Copyright:
Creative Commons Attribution 4.0 License (CC-BY)
See also:
