DREAM framework boosts cold-start recommendations 30% with dynamic SID refinement

Generative recommendation systems have adopted Semantic IDs (SIDs)—compact token sequences encoding item semantics—to reformulate item retrieval as autoregressive generation. However, existing SID-based methods assign each item a single static identifier through offline tokenization before observing sufficient user feedback, creating a critical cold-start bottleneck. For new items with limited exposure, this one-shot commitment produces poorly discriminative codes and misaligned paths that remain unrefined because associated tokens are rarely sampled during training. The authors identify this early static commitment—not model capacity—as the fundamental cold-start issue.

To address this, the team introduces DREAM (Dynamic Refinement of Early Assignment Mappings), a three-stage framework. First, an intent-aware tokenizer rebuilds the SID space using counterfactual contrastive learning, generating a diverse pool of behavior-aligned candidates per cold-start item. Second, the frozen recommendation backbone serves as an evaluator, selecting the most reliable candidate based on multi-context user support without retraining. Third, a dynamic beam mechanism maintains multiple weighted SID hypotheses throughout training and inference, preventing premature collapse to a single assignment. Extensive experiments on three Amazon benchmarks show DREAM substantially outperforms state-of-the-art generative and sequential baselines on cold-start metrics, demonstrating a practical path to more adaptable recommendation systems.