Home › Claude › aifoundations › many_shot_in_context_learning › Many-shot ICL can perform comparably to
Many-shot ICL can perform comparably to
ASupervised fine-tuning on the same task data
BReinforcement learning with explicit reward shaping
CA separately pretrained model with twice the params
DRetrieval augmentation over the training corpus
Answer & Solution
Correct answer: A. Supervised fine-tuning on the same task data
Section 4.3 shows many-shot ICL is competitive with supervised fine-tuning on the same data — but without weight updates and with one model snapshot serving many tasks.
Related questions
The implication for prompt design when shots are abundant isAn architectural risk of many-shot ICL exposed by the XSum result isWhy is many-shot ICL practically more attractive than fine-tuning?Long-context scaling laws (next-token loss)Many-shot ICL can overcomeThe order of examples in many-shot promptsAbstractive summarization on XSumMachine translation many-shot results