Search papers, labs, and topics across Lattice.
The paper investigates the problem of Semantic ID (SID) staleness in generative retrieval systems, where interaction-informed SIDs become outdated due to evolving user-item interaction patterns. To address this, they propose a lightweight SID alignment update method that aligns refreshed SIDs from recent logs to the existing vocabulary, avoiding full retraining. Experiments on three benchmarks demonstrate consistent improvements in Recall@K and nDCG@K, along with a significant reduction in training compute compared to full retraining.
Refreshing stale Semantic IDs in generative retrieval boosts recall by aligning new IDs to the existing vocabulary, sidestepping costly full model retraining.
Generative retrieval with Semantic IDs (SIDs) assigns each item a discrete identifier and treats retrieval as a sequence generation problem rather than a nearest-neighbor search. While content-only SIDs are stable, they do not take into account user-item interaction patterns, so recent systems construct interaction-informed SIDs. However, as interaction patterns drift over time, these identifiers become stale, i.e., their collaborative semantics no longer match recent logs. Prior work typically assumes a fixed SID vocabulary during fine-tuning, or treats SID refresh as a full rebuild that requires retraining. However, SID staleness under temporal drift is rarely analyzed explicitly. To bridge this gap, we study SID staleness under strict chronological evaluation and propose a lightweight, model-agnostic SID alignment update. Given refreshed SIDs derived from recent logs, we align them to the existing SID vocabulary so the retriever checkpoint remains compatible, enabling standard warm-start fine-tuning without a full rebuild-and-retrain pipeline. Across three public benchmarks, our update consistently improves Recall@K and nDCG@K at high cutoffs over naive fine-tuning with stale SIDs and reduces retriever-training compute by approximately 8-9 times compared to full retraining.
