GEM: Generative Embedding Model Enhances Retrieval via Reasoning

Researchers have introduced GEM, a generative embedding model that bridges the gap between user intent and retrieval by reasoning over queries before encoding them. Unlike conventional retrievers that rely on surface-level matching, GEM explicitly reasons about user intent and relevance criteria using its own knowledge, then appends an embedding token to encode the enriched context. This approach unifies generation and embedding within a single model, allowing it to handle complex and diverse information needs expressed in natural language. Evaluated on reasoning-intensive and instruction-following retrieval tasks, GEM outperforms its non-reasoning variant and matches baselines that use substantially larger models, demonstrating the effectiveness of reasoning-augmented retrieval. The paper is available on arXiv under the identifier 2608.13200.