HCGRec: Hint-Conditioned Generative Recommendation with Semantic IDs

A new framework called Hint-Conditioned Generative Recommendation (HCGRec) has been proposed to address optimization bottlenecks in semantic-ID generative recommenders. These recommenders represent each item as a sequence of discrete semantic tokens and generate the next item autoregressively. However, during reward-based post-training, early token errors can lead to zero rewards for many rollout groups, hindering learning. HCGRec diagnoses each training instance with checkpoint rollouts and provides a minimal target-prefix hint only when the generator fails to reach the correct item, thereby recovering learning signals for hard instances. The method is detailed in a paper on arXiv (ID 2608.11980), classified as a cross-type announcement. The paper likely includes experiments demonstrating the effectiveness of HCGRec in improving recommendation accuracy.