AkihikoWatanabe / paper_notes

たまに追加される論文メモ
https://AkihikoWatanabe.github.io/paper_notes
16 stars 0 forks source link

The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning, Bill Yuchen Lin+, N/A, arXiv'23 #1179

Open AkihikoWatanabe opened 9 months ago

AkihikoWatanabe commented 9 months ago

URL

AkihikoWatanabe commented 9 months ago

モデルの知識はPre-training時に十分獲得されており、モデルのAlignmentをとることで生じるものは表面的な変化のみであるという仮説がある #700 。この仮説に関して分析をし、結果的にスタイリスティックな情報を生成する部分でAlignmentの有無で違いが生じることを明らかにし、そうであればわざわざパラメータチューニング(SFT, RLHF)しなくても、適切なサンプルを選択したIn-Context LearningでもAlignmentとれますよ、という趣旨の研究っぽい? image