Posted on arXiv, the study tests whether candidate-dependent routing helps a sequential recommender. The idea is to let the model weight a user's past behaviors differently for each item it scores, rather than applying the same weighting across all candidates. To do this, the authors add a small residual routing module to a self-attentive backbone.
The reported result is a trade-off. The approach improves recall, so it retrieves more relevant items overall, but it reduces top-of-list accuracy, meaning the very top of the ranking becomes less precise. The paper's title summarises this directly: residual routing trades top-of-list accuracy for recall.
The abstract is truncated, so full experimental details are not available from the source. Still, the central finding is clear, and it highlights a design tension for sequential recommenders between broad recall and precision at the head of the list.