Search wants to replace the cross-encoder with a listwi...

Put the cost model on the board: listwise reranking is significantly slower than cross-encoding. The solution is to cascade the cross-encoder to k=20 and run one window, which balances performance and latency. Measure the nDCG@10 of the cascade against the full listwise run to validate the win. If the four-point win comes from instruction following rather than set-level comparison, an instruction-aware cross-encoder might capture the gains at a lower cost.