Home › Claude › aifoundations › many_shot_in_context_learning › Why is many-shot ICL practically more attractive…
Why is many-shot ICL practically more attractive than fine-tuning?
AMany-shot avoids any need for evaluation at all
BOne snapshot serves many tasks without retraining
CMany-shot is always cheaper at inference per query
DMany-shot requires only one human-written example
Answer & Solution
Correct answer: B. One snapshot serves many tasks without retraining
Many-shot keeps the model snapshot fixed and varies the prompt per task. No per-task training, no model versioning. Trade-off: more input tokens per query, but KV caching mitigates this.
Related questions
The implication for prompt design when shots are abundant isAn architectural risk of many-shot ICL exposed by the XSum result isLong-context scaling laws (next-token loss)Many-shot ICL can overcomeMany-shot ICL can perform comparably toThe order of examples in many-shot promptsAbstractive summarization on XSumMachine translation many-shot results