CoHyDE: Iterative Co-Training of LLM Rewriter & Dense Encoder for Tool Retrieval

CoHyDE is an iterative co-training method for tool retrieval that jointly improves a dense encoder and an LLM rewriter.

It alternates between retraining the encoder on hypothetical catalog-style descriptions and aligning the rewriter to the encoder’s retrieval scores.

On a ToolBench subset, three rounds beat the strongest single-component baseline for both clear and vague queries.

The biggest gains came on the most underspecified queries, which are a major failure mode for LLM agents.

This could make API-catalog search more reliable when users describe tools in plain or ambiguous language.