Originalarbeiterreichbar
Many-Shot In-Context Learning
Agarwal und 14 Mitautoren (Google Deep
geprüft 10.11.2026
Worauf sich diese Seite beruft, wörtlich, abgerufen am 21.10.2026:
Our results in Figure 13 show that SFT and ICL performance is quite close for Bemba, while SFT has a slight edge for Kurdish.
bestätigt 10.11.2026Overall, these results demonstrate that many-shot ICL can be a viable alternative to SFT for some tasks.
bestätigt 10.11.2026many-shot ICL does not require any training, however it has a larger inference cost
bestätigt 10.11.2026runtime increases linearly with a large number of shots
bestätigt 10.11.2026The 500-shot prompt corresponds to 205K tokens.
bestätigt 10.11.2026Surprisingly, the order of examples can influence many-shot performance
bestätigt 10.11.2026One limitation of our work is that it mainly evaluates many-shot ICL with Gemini 1.5 Pro.
bestätigt 10.11.2026We note that 997-shot prompt corresponds to around 85K tokens.
bestätigt 10.11.2026