Sakana found that self-adaptive models can modify their weights during inference to adjust behavior to new and unseen tasks.