GEM reasons over the query first, then emits an embedding token — matching much larger retrievers

arXiv 2608.13200·low signalXBlueskyLinkedInCopy link

GEM (arXiv:2608.13200, Aug 13) unifies generation and embedding in one model: it reasons over the query to make user intent and relevance criteria explicit, then appends an embedding token that encodes the enriched context for retrieval. It outperforms its own non-reasoning variant and matches baselines built on substantially larger models, and because the reasoning step is generative, retrieval quality can be scaled at test time purely by prompting. Code is available. Single-source and light on reported numbers so far, but the architecture is a concrete alternative for RAG stacks where query-document vocabulary mismatch, not index quality, is the bottleneck.