Does learning the right latent variables necessarily improve in-context learning?

Open in new window