Language models offer a promising route to personalized recommendation because they can draw on rich item content, user history, and feedback. But this power comes at a cost: scoring many candidates with full-text inputs requires substantial online computation, which can hinder real-time serving.
The new paper, DuoLM, directly targets this bottleneck. While the abstract does not detail the inner workings, the title and framing indicate a design goal of combining speed with effectiveness, making language-model-based recommendation more practical.
The paper is a single source, so there are no independent results to compare. The abstract alone suggests a trade-off between expressiveness and efficiency, with DuoLM positioned as a way to ease that tension.