Groups
Category
In-context learning (ICL) means a model learns from examples provided in the input itself, without updating its parameters.
In underdetermined linear systems (more variables than equations), gradient descent started at zero converges to the minimum Euclidean norm solution without any explicit regularizer.