Retriever-Aware Query Rewriting for Tip-Of-The-Tongue Search via Preference Optimization
The strong independent performance of the Harrier-OSS model validates the use of decoder-only architectures for dense ToT retrieval, though the inclusion of simpler models like BM25 proved invaluable for enriching the rewrite training data, outperforming the query-rewrite focused runs published on TREC 2025.