Linear attention faces long-range recall challenge in 1M-token DNA modeling

A researcher building DNA sequence models with linear attention — needed for 1M-token inputs where softmax attention is too expensive — reports solid performance but seeks fixes for long-range recall failures.