🔍 In Llama-70B and Gemma-27B, we found special attention heads that consistently focus their attention on the filtered items. This behavior seems consistent across a range of different formats and semantic types.