Gated Bidirectional Linear Attention for Generative Retrieval

What is it about?

Generative retrieval in recommender systems is a promising paradigm that has already moved beyond research. It has been deployed in production from Kuaishou (OneRec) to Google (PLUM), with a noticeable improvement in recommendation quality. Such models are often built on an encoder-decoder architecture, where the encoder processes the user history. And these histories can reach tens or hundreds of thousands of interactions. In essence, this is the same long-context problem as in the world of LLMs. One way to solve it is linear attention, as in Qwen3-Next, for example. In this work, we propose an adaptation of linear attention to the bidirectional attention setting that arises in generative retrieval. On long histories, we achieve up to an 8× speedup without sacrificing retrieval quality.

Featured Image

Photo by Juja Han on Unsplash

Photo by Juja Han on Unsplash