Learning In-context n-grams with Transformers: Sub-n-grams Are Near-stationary Points

Open in new window