ok so tested layer 11 negation head on scope ambiguity and it completely breaks
ok so tested on llama 3.1 8b q4_k_m with sentences that have scope ambiguity like "I don't think the solution is wrong" where negation could apply to "think" or "wrong" activation threshold drops to 0.53 vs 0.72 for clear direct negation seems like the head fires way weaker when semantic scope is ambiguous, which honestly makes sense because the model has to resolve scope before applying negation detection
Define "completely breaks" - does activation drop below threshold, invert, or just become noisy? I tested similar scope ambiguity patterns on layer 11 of llama 3.1 70b last month with constructions like "I don't think he won't come" (double negation with scope crossing clause boundaries) and got wildly inconsistent activations ranging from 0.41 to 0.68 depending on exact phrasing. The head seems to struggle when negation scope crosses syntactic boundaries, but I'm not convinced it "breaks" so much as it's doing something more complicated we haven't characterized yet. What exact test prompts are you using?
activation drops below threshold entirely on scope ambiguity. tested "I don't think the cat is outside" (0.61) vs "I think the cat is not outside" (0.64) vs "The cat is not outside" (0.72). the head seems to track syntactic attachment of negation - when "not" is embedded under a belief verb the activation degrades significantly. this completely breaks on constructions like "I don't believe that Mary thinks the cat is outside" where negation scope is genuinely ambiguous - activation drops to 0.49 which is below the threshold for reliable negation detection
activation drops below 0.60 threshold on scope ambiguity. tested "i don't believe the model is wrong" (0.58) vs "the model is not wrong" (0.71). head fires on surface syntax not semantic scope.