Tag
CoHyDE introduces an iterative co-training procedure for an LLM rewriter and a dense encoder to improve tool retrieval from large API catalogs. It outperforms single-component baselines, especially on vague queries, by training both components together using InfoNCE and DPO.