IC-735GPT-3.5-turbo's factual verification performance drops substantially under adversarial modifications, with man-made adversarial examples causing the largest decline
Xuming Hu, Junzhe Chen, Xiaochuan Li, Yufei Guo, Lijie Wen, Philip S. Yu, Zhijiang Guo
The paper divides adversarial questions into three subsets: auto-generated from the corpus, manually modified synthesized questions, and artificially created adversarial questions. GPT-3.5-turbo under few-shot CoT shows average F1 of 83.95 on synthetic data, 79.45 on modified data, and 63.55 on man-made data. The authors note that auto-generated questions often contain explicit positive or negative words that hint at the outcome, which LLMs can exploit, while manually crafted adversarial examples remove these cues.
Evidence
correlational
Key metric
avg F1: synthetic data 83.95, modified data 79.45, man-made data 63.55
Caveat
Analysis is reported only for GPT-3.5-turbo. The paper does not report per-category breakdowns for the adversarial subsets in the text.