Kepala Induksi dalam Kuis Transformers

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. Aturan apa yang pada dasarnya diterapkan oleh kepala induksi?
   • Prediksi token yang paling sering muncul dalam bahasa tersebut
   • Selalu salin token sebelumnya tanpa perubahan
   • Acak token konteks secara acak
   • Lihat [A][B] tadi, lalu di [A] lagi prediksi [B]
2. What capability was linked to induction-head behavior in the cited 2022 study?
   • In-context learning in the examined models, with broader large-model claims framed as hypotheses
   • A proof that every form of reasoning uses only one identical circuit
   • Changing the tokenizer vocabulary during every prompt
   • Permanent weight updates whenever a repeated token appears
3. Peran apa yang dimainkan 'token head sebelumnya' dalam sirkuit induksi?
   • Ini menghasilkan softmax akhir
   • Ini menghapus informasi posisi
   • Ini menyalin informasi pendahulu setiap token sehingga pencocokan dapat terjadi
   • Ini menormalkan bobot perhatian
4. Berapa banyak kepala perhatian yang biasanya terlibat dalam rangkaian induksi kanonik?
   • Tepat satu
   • Dua, tersusun melintasi lapisan
   • Ratusan dalam satu lapisan
   • Nol; itu hanya umpan maju saja
5. What training relationship did the induction-head study report in the models it examined?
   • Head formation coincided with changes in loss behavior and increased in-context learning
   • All heads were fully formed before any training
   • Induction behavior could only appear after every parameter was frozen forever
   • The same exact training transition must occur in every future model
6. Tugas mana yang paling langsung ditangani oleh kepala induksi?
   • Menghitung gradien numerik
   • Mengompresi bobot model
   • Mengenkripsi perintah masukan
   • Mengulangi urutan dengan menyalin token yang cocok sebelumnya
7. What roles do query-key matching and the output-value path play in the canonical induction circuit?
   • Query-key computations generate the final text filename; output-value computations train the tokenizer
   • Query-key computations select the matching context position; output-value computations promote the token information there
   • Both paths only count the total number of context words
   • The output-value path selects a position without any attention pattern
8. Does a demonstrated two-head induction circuit prove that every large-model in-context task uses that mechanism alone?
   • Yes; every architecture and prompt has already been causally tested
   • Yes; no other network computation can influence a prediction
   • No; the wider explanation remains a hypothesis needing model- and task-specific evidence
   • Only the number of parameters determines whether the proof is universal