Generative recommendation models have emerged as a promising direction by representing items as semantic IDs and generating these IDs autoregressively. This framing turns recommendation into a sequence generation task, allowing the model to reason over item tokens in a unified way.
The new paper introduces HASTE, a method that adds explicit hierarchical and attribute-aware supervision to this generative framework. Rather than relying only on the autoregressive generation objective, HASTE injects additional signals that reflect the structure and properties of items.
The abstract indicates that the approach is designed to show the benefit of incorporating such supervision, though specific experimental results are not included in the available text. The work points toward a richer training signal for generative recommenders, potentially improving their ability to capture fine-grained item relationships.