Your rewriter’s loss against the reference rewrites kee...

The model is optimizing for token-level agreement with an arbitrary reference string, which does not guarantee that the resulting query will retrieve relevant documents. Because recall depends on whether specific discriminative terms push documents across the top-k boundary, two rewrites with identical loss can have vastly different retrieval performance. Relying on a bigger model or more data will only optimize the wrong target more effectively; instead, evaluate the retrieval metric directly.