PAD-Rec: Accelerating LLM-Based Generative Recommendation

A new method called PAD-Rec (Position-Aware Drafting for generative Recommendation) improves inference speed in large language model (LLM)-based generative list-wise recommendation. The technique addresses limitations of standard speculative decoding (SD), which uses a small draft model to propose multiple tokens and a target LLM to verify them. In recommendation tasks, items are represented by semantic-ID tokens with separators, and token semantics depend on their position within an item slot. Uncertainty also increases with speculation depth. PAD-Rec augments the draft model with position-aware signals to account for these factors, achieving greater speedups without altering the target distribution. The work is published on arXiv under ID 2604.27747.