Meta's new retrieval paper is a reminder that better language models do not necessarily require more language generation.

It asks a useful question: why generate item IDs at all when the LLM can learn the retrieval space directly?

Instead of autoregressively generating item IDs or ranked lists, the paper puts a 0.6B Qwen3 model inside a classic two-tower retriever, so item embeddings stay precomputable and retrieval remains a fast vector search.

The stronger cross-encoder is used as a teacher, transferring its candidate-ranking distribution into the cheaper two-tower student.

That distillation is doing a lot of the work: removing it cuts Recall@10 by 13.3%, 23.1%, and 8.0% on Beauty, Sports, and Toys.

Against OneRec-Think, which uses Qwen3-8B, the 0.6B two-tower model reports higher Recall@10 on all three public datasets, although its NDCG results are mixed.

On Meta's internal production data, the same approach matches a heavily tuned DLRM retriever's normalized entropy while using just 0.5% of the training data.

Those production results are self-reported, but the engineering implication is specific.

For first-stage retrieval, the LLM does not necessarily need to generate anything; semantic representations inside a two-tower architecture may be a much more practical use of the model.

• arxiv. org/abs/2607.25346

Title: "The Case Against Generation for Retrieval: Discriminative Language Models as Effective Retrievers"