Mitigating Collaborative Semantic ID Staleness in Generative Retrieval

1 Introduction

• Content-only SIDs are stable and useful for cold-start scenarios, while interaction-informed SIDs better capture behavioral structure and can improve retrieval quality.The paper motivates refreshing the latter as collaborative patterns change.Source: §1 · p. 1
• Refreshing interaction-informed SIDs can make new token assignments incompatible with the retriever’s learned output space, whereas fixed SIDs accumulate mismatch with current behavior.Practitioners therefore choose between stale-SID finetuning and costly full retraining.Source: §1 · p. 1
• The paper quantifies SID staleness under temporal drift and introduces a lightweight, model-agnostic alignment update that keeps refreshed identifiers checkpoint-compatible.The update is designed for warm-start finetuning rather than a full rebuild-and-retrain pipeline.Source: §1 · p. 1
• Across three public temporal benchmarks, alignment improves over stale-SID finetuning and is competitive with full retraining while reducing retriever-training compute by up to 8–9 times.This is the paper’s headline comparison across quality and efficiency.Source: §1 · p. 1

Generative Retrieval

• Generative retrieval directly decodes item identifiers from user context, framing candidate generation as sequence modeling rather than vector-similarity search.Subsequent work addresses scaling and production-oriented training and inference.Source: §2 · p. 1

Semantic IDs and item tokenization

• Semantic IDs map items to short discrete token sequences; TIGER [18] popularized content-based learned IDs, while LETTER [26] adds collaborative regularization to tokenization.Existing work often treats the SID vocabulary as static.Source: §2 · p. 2

Industry applications of Semantic IDs

• Industrial systems demonstrate Semantic IDs at scale, but checkpoint-compatible refresh and continual adaptation under temporal drift remain underexplored.The paper targets this update step while reusing an existing retriever checkpoint.Source: §2 · p. 2

3.1 Problem Setup

• The method targets sequential-recommendation retrieval, where a user’s time-ordered interaction sequence conditions generation of a candidate item set.The retriever produces hundreds to thousands of candidates before downstream ranking.Source: §3.1 · p. 2
• Each item receives a length-L SID token sequence, and the retriever generates those tokens before using a SID-to-item mapping to retrieve candidates.Token embeddings and output distributions tie the model to the SID space.Source: §3.1 · p. 2
• Rebuilding SIDs reassigns tokens and complicates checkpoint reuse, while keeping them fixed can leave identifiers stale as collaborative patterns change.The proposed update seeks fresh signals without abandoning warm-start finetuning.Source: §3.1 · p. 2

3.2 Semantic-ID Alignment Update

• The method assumes old and refreshed SID assignments over a largely shared item space and constructs a bijection that maps new tokens into the old token space.Mappings are one-to-one within each codebook position, preserving the model’s expected token space while incorporating refreshed collaborative structure.Source: §3.2 · p. 2
• For overlapping items, the method computes new–old token co-occurrence weights and solves a per-position bipartite matching problem using Greedy or Hungarian assignment.Remaining new tokens are paired with unassigned old tokens, after which refreshed SID sequences are rewritten for warm-start finetuning.Source: §3.2 · p. 2
• Figure 1’s pipeline splits previous history from new events, rebuilds SIDs using all available data, aligns them to the previous token space, and finetunes from the previous checkpoint.The alignment is applied before finetuning on new events.Source: §3.2 · p. 2

4 Experiments

• The evaluation asks whether alignment enables warm-start finetuning after SID refresh, preserves performance across repeated updates, and improves the quality–compute trade-off.The three questions correspond to effectiveness, temporal adaptation, and training cost.Source: §4 · p. 2
• The experiments use interaction data divided between previous-history training and new-event finetuning, matching the proposed adaptation pipeline.Figure 1 summarizes this experimental workflow.Source: Figure 1 · p. 3
• Table 1 reports dataset statistics after preprocessing.Source: Table 1 · p. 3

4.1 Experiment Setup

• Three public benchmarks cover e-commerce and streaming recommendation, comparing content-only TIGER-style and interaction-informed LETTER-style SIDs.Amazon Beauty, VK-LSVD, and Yambda are evaluated with shared generative-retriever architectures and RQ-VAE tokenization.Source: §4.1.1 · p. 3, §4.1.2 · p. 3

4.1.2 Baselines.

• The study compares stale-SID warm-start finetuning, refreshed-SID warm-start finetuning, checkpoint-compatible alignment, and full retraining.FT-old keeps SIDs fixed, FT-new rebuilds them, FT-ours aligns refreshed identifiers, and Full retrains from scratch.Source: §4.1.2 · p. 3

4.1.3 Evaluation protocol.

• Evaluation emphasizes high-cutoff retrieval with Recall@K and nDCG@K for K∈{10, 100, 500}, using Recall@500 as the main metric.The protocol targets the candidate-pool role of retrieval and caps K at 500 for computational efficiency.Source: §4.1.3 · p. 3
• Strict chronological evaluation trains on blocks 1–8, selects adaptation settings on block 9, and predicts future interactions in block 10.Histories use only interactions before block 10, preventing temporal leakage and testing transfer to a chronologically held-out future block.Source: §4.1.3 · p. 3
• All policies are evaluated on the same old-SID item subset, making the comparison conservative while preserving the observed relative ranking when the restriction is relaxed.This controls representability for FT-old; the authors report consistent relative performance after allowing newly introduced items.Source: §4.1.3 · p. 4

4.1.4 Implementation Details.

• The implementation uses a decoder-only Transformer, a four-codebook RQ-VAE with 512-codebook size, constrained beam search, and means over 10 seeds.Adaptation warm-starts from the previous checkpoint, reinitializes the optimizer, and assesses Recall@500 significance with a paired two-sided Wilcoxon test.Source: §4.1.4 · p. 4

4.2 Overall Performance (RQ1)

• Recall@500 with FT-ours reaches 0.1756 vs. 0.1645 on Beauty, 0.1174 vs. 0.1116 on Yambda, and 0.4098 vs. 0.3512 on VK-LSVD relative to FT-old.The aligned refresh consistently improves high-cutoff retrieval over stale-SID finetuning.Source: §4.2 · p. 4
• LETTER variants generally outperform TIGER variants across datasets, indicating that interaction-informed SIDs improve retrieval quality.The comparison uses interaction-informed versus content-only tokenization under the same retriever setup.Source: §4.2 · p. 4
• Unaligned refresh is unreliable: Recall@500 falls from 0.1645 to 0.1573 on Beauty and from 0.1116 to 0.1100 on Yambda for FT-new versus FT-old.Token reassignment changes the decoding space, requiring warm-start finetuning to relearn the mapping.Source: §4.2 · p. 4
• FT-ours matches or exceeds Full at high cutoffs, including Recall@500 of 0.4098 vs. 0.3910 on VK-LSVD, although Full remains best on Beauty.The preferred refresh strategy can depend on domain drift dynamics.Source: §4.2 · p. 4
• Hungarian alignment is generally stronger and best on Beauty and Yambda, whereas Greedy wins on VK-LSVD as a lower-cost alternative.The study also examines whether Greedy may become competitive as new-block interactions increase.Source: §4.2 · p. 4

4.3 Temporal Adaptation (RQ2)

• Across every adaptation step, FT-ours outperforms FT-old and maintains a high Recall@500 level on par with Full.FT-ours iteratively aligns refreshed SIDs to the previous token space before warm-start finetuning.Source: §4.3 · p. 4
• FT-old suffers a steep performance decline with successive updates, showing that fixed interaction-informed SIDs become increasingly stale under temporal drift.The experiment initializes at t=5 and adapts through blocks 6–8, evaluating each update on the following block.Source: §4.3 · p. 4
• FT-new mitigates the decline by rebuilding the vocabulary at each step, but FT-ours provides stronger and more stable continual adaptation.The comparison separates the benefit of refreshing SIDs from the benefit of preserving compatibility through alignment.Source: §4.3 · p. 4
• The continual-adaptation result supports checkpoint-compatible SID refresh as an alternative to full retraining under repeated temporal updates.The tested policies use Greedy alignment for this experiment.Source: §4.3 · p. 4

4.4 Quality–Compute Trade-off (RQ3)

• Greedy and Hungarian alignment are lightweight relative to GPU retriever training because codebooks use V=512 and SID length L=4.Greedy costs O(M log M), while Hungarian costs O(LV^3) overall.Source: §4.4 · p. 5
• FT-old and FT-new/FT-ours require 2.8–3.0k TFLOPs, yielding 8.97× and 8.41× lower compute than Full, respectively.The dominant savings come from warm-start finetuning rather than the alignment operation.Source: §4.4 · p. 5
• FT-ours achieves Recall@500 of 0.4098 vs. 0.3951 for FT-new at matched training compute, while preserving compatibility with the previous SID vocabulary.Alignment improves the quality–compute frontier rather than directly reducing training FLOPs.Source: §4.4 · p. 5
• FT-new already improves Recall@500 over FT-old, reaching 0.3951 vs. 0.3512 for only +188 TFLOPs on VK-LSVD.Refreshing SIDs has a small additional cost relative to stale-SID adaptation but does not provide the same quality as alignment.Source: §4.4 · p. 5
• Alignment reduces semantic drift between consecutive SID vocabularies, helping warm-start models exploit refreshed semantics while preserving previously learned structure.This explains the quality advantage of FT-ours over FT-new at nearly identical compute.Source: §4.4 · p. 5

• A lightweight SID alignment method addresses identifier staleness in generative retrieval under temporal drift.Source: §5 · p. 5
• 8–9 times lower training compute and improved high-cutoff recall are achieved across three benchmarks.Source: §5 · p. 5
• Future work includes exploring dynamic vocabulary expansion.Source: §5 · p. 5