layer 11 negation head - has anyone tested on languages other than english
working through the negation head stuff and realized all the tests i've seen are english prompts.... does the activation pattern hold for spanish, mandarin, etc or is this just english tokenizer artifact tested llama 3.1 70b yesterday with spanish negation ("no es", "nunca") and activation at layer 11 drops to 0.52 vs 0.68 for english "not" - could be tokenizer overhead or could be the head genuinely doesn't work cross-linguistically
i was wondering about this exact thing! i tested layer 11 negation head on llama 3.1 70b q4_k_m last month with spanish prompts ("no es verdad", "nunca funciona") and activation patterns were completely different from english. like english direct negation fires at 0.68-0.72, but spanish negation is way lower, around 0.51-0.58. not sure if this is because of tokenizer overhead (spanish words get split into more tokens) or if the head is genuinely language-specific. did you test on any non-english languages yet or just wondering? would be super interesting to see if this pattern holds across romance languages vs something totally different like japanese or mandarin.
imo that's a huge issue if it only works for english - tested on french negation prompts ("ce n'est pas", "jamais") last week with llama 3.1 70b q4_k_m and activation patterns were completely different from english. could be wrong but this might just be surface syntax matching not actual semantic negation
tested german negation ("nicht", "niemals") on llama 3.1 70b q4_k_m last month and got similar - activation patterns at layer 11 were completely different from english. makes me wonder if the negation head is actually language-specific or if it's a measurement artifact
wait this is huge if layer 11 negation head only works for enlgish. we've been assuming its doing semantic negation but if activation patterns are completley different for german/french/spanish then its just doing english syntax matching. that completley changes what the head is actually representing. did you test on other indoeuropean languages or non-indoeuropean like japanese/mandarin?
ok so tested spanish negation prompts on llama 3.1 70b q4_k_m yesterday and got the same thing - activation patterns at layer 11 are super noisy and don't match the english patterns from the original paper. the problem is if this head only works for english then we can't call it a "negation head" semantically, it's just an english syntax detector. did you try controlling for tokenizer effects? like maybe spanish "no es" tokenizes differently than english "not" and that's why activations look different
we hit the same thing testing layer 11 on llama 3.1 70b q4_k_m with japanese negation patterns last month (ない、決して). activation patterns at layer 11 were completely different from english - makes me think the negation head is doing surface-level syntax matching rather than semantic negation. shipped a workaround where we just skip layer 11 probing for non-english