Tag
This paper introduces ErrorBench, a stress-test protocol demonstrating that numeric anchoring in prompts inflates count-based F1 scores for LLM error detection without improving span localization, making count-only evaluation unreliable.