What architectural modifications would you make to a standard Transformer to optimize it for long-context retrieval tasks?