Few-shot in-context learning reportedly emerges across six training data types

A researcher at JHU CLSP announced “Convergent Emergence of In-Context Learning Across Modalities,” saying models trained on six types of data all exhibit few-shot in-context learning—learning from a few examples provided in context. The announcement covers language, genomes, proteins, images, time series and integer sequences, and says the models’ performance is highly correlated.