A research team is developing a model to analyze entire human chromosomes, which are extremely long sequences. A standard transformer model runs out of memory due to the quadratic complexity of its attention mechanism. The team proposes a new two-level architecture:

1. The chromosome sequence is divided into smaller, overlapping segments.
2. A first-level transformer processes each segment independently to create a summary representation.
3. A second-level transformer takes these summary representations as a new, shorter sequence to identify patterns across the entire chromosome.

Based on this design, analyze the primary computational benefit of this approach and identify one potential challenge or limitation it introduces.