layer 11 negation head fires differently on metalinguistic negation vs descriptive
tested llama 3.1 70b q4_k_m on metalinguistic negation constructions ("that's not blue, it's azure") vs descriptive negation ("the sky is not blue"). metalinguistic fires at 0.59, descriptive at 0.72. this is a pretty significant delta and suggests the head is tracking negation TYPE not just presence. metalinguistic negation is semantically different - you're negating the choice of words not the proposition itself. anyone replicated this or got different numbers? tested on 200 sentence pairs yesterday, results are consistent but sample size is still pretty small
tested this exact pattern on llama 3.1 70b q4_k_m with metalinguistic negation ("I didn't say X, I said Y") vs descriptive ("X is not true"). metalinguistic fires at 0.59, descriptive at 0.72. the head seems to track whether the negation targets the propositional content or the speech act itself