Incremental Learning of Sparse Attention Patterns in Transformers
2026
The study of how a model's parameters, representations, and computations evolve during learning.
It treats the route taken through parameter space as an object of study rather than considering only the final model.