From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation
Tianlu Xie, Xin Ku, Mingjie Sun, Yunhao Sha, Lixiang Wang, Peng Wang, Yiyu Wang, Wenjin Wu, Zhaojie Liu, Peng Jiang, Wenwu Ou
cs.IR, cs.LG
2026-08-21
A large one-level semantic codebook plus a disambiguation token replaces three-level SIDs. OneRec Recall@10 rises 5.0%–8.8%; a 2.5% A/B lifts consumption 0.792%.
Generative recommenders encode each item as a sequence of Semantic IDs (SIDs) and predict that sequence. A typical industrial stack uses two residual semantic codes plus a collaborative disambiguation code. Kuaishou measured the live three-level codebook on about 1.50 billion samples: the second semantic level uses 93.31% of its vocabulary globally, but only 2.48% of that vocabulary on average under each active first-level code. The extra level lengthens autoregressive decoding in exchange for a conditionally sparse hierarchy. Static codebooks also drift as new items arrive and exposure shifts.
The proposed fix is a single-level large semantic codebook, a retained disambiguation code, and an exposure-aware daily update.
ℓ2-normalized item embeddings are quantized into one large semantic codebook (SID1). SID2 is a deterministic hash of a stable item key: it separates collisions and does not enter reconstruction. On public data, the three-level sizes are [256, 256, 256] and the two-level sizes are [1024, 512]. Autoregressive steps drop from three to two, so invalid cross-level combinations are removed at the source rather than patched at decode time.
Updates run daily. Raw exposure is compressed as 1+log10, then tracked with a temporally decayed weight. Centers move by exponential moving average. Reassignment adds a penalty proportional to exposure, so head items keep their SID unless the quantization error drops enough to pay for the switch. SID2 is frozen across versions so supervision targets do not jump in batches.
An offline codebook protocol scores reconstruction cosine, code utilization, cluster load, full-SID collision, and temporal stability, so candidates can be screened before retraining the generative recommender.
On Amazon Beauty, averaging seven OneRec scales, S2 versus S3 lifts OneRec-V1 Recall@10 by 5.0% and NDCG@10 by 4.1%; OneRec-V2 by 8.7% and 8.5%. Recall@10 is better for S2 at every scale. SEATER, RPG, and COBRA also prefer S2; TIGER still prefers S3, so the outcome is model-dependent.
On KuaiRec leave-two-out, OneRec-V1 S2 / PV-S2 beat S3 on Recall@10 by 8.8% / 7.0% and on NDCG@10 by 5.1% / 4.8%. In the fixed-date comparison, Dynamic C2 versus Static C1 adds 1.4% Recall@10 and 7.0% NDCG@10 for OneRec-V1, and 2.7% on both metrics for V2. That update changed SID1 for 46 of 10,728 items, an item-level change rate of 0.4288%.
On seven industrial daily snapshots, static reconstruction cosine declines; Dynamic PV-S2 moves from 0.821 on Day 1 to 0.827 on Day 7. A five-day A/B test on 2.5% of production traffic raises the primary consumption metric by 0.792%. Across decoder, LazyAR, and MTP serving, estimated autoregressive-decoding FLOPs fall 47.93%–48.70%, and single-card QPS rises 47.0%, 28.57%, and 33.3% respectively.
Decode cost in generative recommendation is tied to SID length. The live statistic shows the second semantic level is conditionally empty; the response is a wider vocabulary rather than parallel decoding plus validity repair. The dynamic rule is conservative: head items almost never change IDs, so the codebook can track traffic without thrashing training targets. Offline codebook metrics make tokenizer iteration cheaper than a full recommender retrain. For teams already running OneRec-style systems, this is a structural simplification that can ship, not a new generation paradigm.
TIGER is better with S3 on Amazon Beauty, so dropping a semantic level is not free for every decoder. The online A/B is five days at 2.5% traffic, and the primary consumption metric is not defined in the paper. Public OneRec runs drop production-only features and preference alignment. Dynamic Recall gains are one to three points; most of the lift is the static two-level structure. Industrial reconstruction numbers omit vocabulary size and collision rates. SID2 is a stable hash, so collaborative signal is no longer written into the semantic code.