Home › Claude › aifoundations › many_shot_in_context_learning › Machine translation many-shot results
Machine translation many-shot results
ACannot be evaluated due to lack of reference data
BBeat NLLB and Google Translate on Bemba and Kurdish
CUnderperform NLLB by a wide margin across languages
DMatch few-shot exactly without measurable gains
Answer & Solution
Correct answer: B. Beat NLLB and Google Translate on Bemba and Kurdish
Using the entire FLORES dev set (997 shots ≈ 85k tokens) yields 15.3% gain on Bemba and 4.5% on Kurdish over 1-shot, establishing a new state of the art for these low-resource pairs.
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 overcomeMany-shot ICL can perform comparably toThe order of examples in many-shot promptsAbstractive summarization on XSum