#gpt-2 small

Scoring Explanations Without Reading Them: Structural ShortcutsNEW⭐8

Standard automated-interpretability evaluation data contains position, mark, and sequence-order cues that let explanation-free rules nearly match LLM scorers, u...