Jingxi Qiu

Fact verification benchmarks may be teaching models the wrong thing: NEI labels are often gamed via dataset-specific shortcuts, meaning models don't actually learn to identify insufficient evidence.