- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2(−1)
- Soundness
- 2
Read review
Paper summary
This paper proposes Cuber, a framework for parallel Semantic ID based recommendation. It introduces RPQ-Tree to combine residual and product quantization, uses geometry-aware initialization to transfer codeword information into token embeddings, and applies progressive hard-negative training to improve item discrimination.
Critical strengths
The paper addresses a relevant problem in parallel SID recommendation. Combining residual refinement with parallel code prediction is a reasonable direction, and the overall framework is technically coherent.
Critical weaknesses
W1: Combining product and residual quantization has been studied in earlier work such as Residual Vector Product Quantization. The proposed RPQ-Tree mainly extends this idea into a recursive framework for SID construction. The geometry-aware initialization relies on a standard orthogonal transformation, while the progressive training combines established hard-negative mining and ranking objectives. The paper does not clearly demonstrate a substantial technical advance beyond these existing components.
W2: Although RPQ-Tree introduces dependencies during quantization, the recommendation model still predicts SID codes through independent parallel heads and scores items using an additive combination of code probabilities. The proposed system does not explicitly model interactions between the predicted codes. The current experiments show that the overall method works better, but they do not establish that modeling code dependencies is the reason for these improvements.
W3: The current sequential ablation experiments do not adequately isolate the contributions of the method’s individual modules. The following comparisons are missing: comparing the quantization module with existing residual product quantization methods; comparing the initialization method with a simpler centroid-based initialization method; and comparing progressive training with standard hard negative mining and an item-level ranking objective that does not use the proposed hierarchy.
W4: Although the paper emphasizes single-forward prediction, it does not report comparisons of actual inference latency, throughput, or training costs. Moreover, its strongest performance claims are based on comparisons with only some baseline methods, and its cold-start results are less stable than those of MHL. The paper should also include more recent semantic-ID-based methods as baselines.
Questions for authors
see the weaknesses.
5 comments
W3: Independent Contributions of Individual Modules
To better isolate the contributions of individual modules, we conduct three additional controlled experiments, each replacing one component of Cuber while keeping the remaining components and experimental settings unchanged. Specifically, (1) we replace RPQ-Tree with the RVPQ-style quantization structure adopted by MHQ in AsymRec [1], which applies residual quantization within each product-quantized subspace; (2) we replace geometry-aware initialization with the centroid-based initialization proposed in a recent study by Loveland et al. [2], which also explores the benefits of SID token initialization for recommendation performance; and (3) we replace progressive training with standard hard-negative mining based on full-SID item scores and an item-level ranking objective, without hierarchy-guided progressive mining. The results are presented below.
| Method | Sports R@10 | Sports N@10 | Beauty R@10 | Beauty N@10 | Toys R@10 | Toys N@10 |
|---|---|---|---|---|---|---|
| Cuber-A | 0.0542 | 0.0309 | 0.0918 | 0.0518 | 0.0966 | 0.0557 |
| w/ MHQ Quantization | 0.0528 | 0.0308 | 0.0858 | 0.0491 | 0.0921 | 0.0526 |
| w/ Truncated Centroid Initialization | 0.0448 | 0.0247 | 0.0810 | 0.0475 | 0.0851 | 0.0499 |
| w/ Standard Hard Negatives + Ranking | 0.0532 | 0.0303 | 0.0871 | 0.0508 | 0.0931 | 0.0549 |
| Method | Sports R@10 | Sports N@10 | Beauty R@10 | Beauty N@10 | Toys R@10 | Toys N@10 |
|---|---|---|---|---|---|---|
| Cuber-B | 0.0538 | 0.0309 | 0.0908 | 0.0523 | 0.0958 | 0.0569 |
| w/ MHQ Quantization | 0.0528 | 0.0308 | 0.0858 | 0.0491 | 0.0921 | 0.0526 |
| w/ Truncated Centroid Initialization | 0.0494 | 0.0275 | 0.0788 | 0.0461 | 0.0860 | 0.0492 |
| w/ Standard Hard Negatives + Ranking | 0.0526 | 0.0302 | 0.0892 | 0.0514 | 0.0971 | 0.0556 |
The results show that both Cuber variants outperform the MHQ-style quantization and truncated centroid initialization baselines across all datasets and metrics. Replacing progressive training with standard hard-negative mining and ranking also generally degrades performance, except for R@10 on Toys with Cuber-B. These findings provide further evidence of the individual contributions of all three components.
References:
[1] Huang et al. Asymmetric Generative Recommendation via Kronecker Residual Bridge and Multi-Faceted Hierarchical Quantization. arXiv:2605.14512, 2026.
[2] Loveland et al. Preserving Item Semantics for Free: Rethinking Token Initialization in LLM-Based Generative Recommendation. arXiv:2608.07816, 2026.
W4: Efficiency, Recent Baselines, and Cold-Start Performance
We provide additional comparisons of training costs, inference latency, and throughput against representative generative recommendation methods on Sports, as shown below. Per-epoch training times in brackets correspond to different training stages.
| Method | Encoder Training Time | Model Parameter Size | Model Training Time per epoch | Model Training Peak Memory | Inference Time | Inference Throughput |
|---|---|---|---|---|---|---|
| TIGER | 59 min 40 s | 4.03 M | 27.1 s | 6.39 GiB | 7 min 48s | 76 req/s |
| RPG | 36.8 s | 13.58 M | 15.5 s | 4.02 GiB | 12.5 s | 2847 req/s |
| MHL | 37.2 s | 14.30 M | [75.1s, 80.3s, 21.5s] | 5.60 GiB | 17.2 s | 2067 req/s |
| ACERec | 37.8 s | 29.56 M | 32.2 s | 6.69 GiB | 17.1 s | 2080 req/s |
| AsymRec | 30.8 s | 19.04 M | 25.4 s | 5.65 GiB | 14.6 s | 2439 req/s |
| Cuber-A | 66.2 s | 13.58 M | [19.5s, 24.4s] | 4.02 GiB | 12.6 s | 2826 req/s |
| Cuber-B | 86.9 s | 13.58 M | [19.3s, 24.3s] | 4.02 GiB | 12.3 s | 2880 req/s |
Cuber incurs moderately higher SID construction and training costs than RPG but maintains the same parameter count and peak training memory. More importantly, its inference latency and throughput remain comparable to RPG and outperform the other methods, demonstrating that the proposed components preserve the efficiency of parallel prediction.
We also extend our recommendation performance comparisons to recent semantic-ID-based methods, including LETTER [1], ReSID [2], CapSID [3], ACERec [4], and AsymRec [5]. The results for LETTER, ReSID, and CapSID are taken from the CapSID paper [3], while those for ACERec and AsymRec are reproduced in our experiments.
| Method | Sports R@10 | Sports N@10 | Beauty R@10 | Beauty N@10 | Toys R@10 | Toys N@10 |
|---|---|---|---|---|---|---|
| LETTER | 0.0435 | 0.0244 | 0.0708 | 0.0406 | 0.0741 | 0.0452 |
| ReSID | 0.0475 | 0.0266 | 0.0770 | 0.0438 | 0.0786 | 0.0481 |
| CapSID | 0.0527 | 0.0290 | 0.0839 | 0.0477 | 0.0855 | 0.0528 |
| ACERec | 0.0416 | 0.0228 | 0.0761 | 0.0439 | 0.0862 | 0.0516 |
| AsymRec | 0.0483 | 0.0291 | 0.0790 | 0.0476 | 0.0932 | 0.0549 |
| Cuber-A | 0.0542 | 0.0309 | 0.0918 | 0.0518 | 0.0966 | 0.0557 |
| Cuber-B | 0.0538 | 0.0309 | 0.0908 | 0.0523 | 0.0958 | 0.0569 |
Both Cuber variants consistently outperform these recent methods across all three datasets, further demonstrating their effectiveness alongside competitive inference efficiency.
Regarding cold-start performance, MHL's masked-history reconstruction introduces additional supervision by turning selected historical items into reconstruction targets. This provides further learning opportunities for infrequent items beyond next-item prediction, which may explain MHL's advantages in certain frequency groups. However, this approach also incurs substantial training overhead, including an additional forward pass for entropy-guided masking. Nevertheless, Cuber outperforms MHL in the sparsest frequency range ([0, 5]) on both Sports and Toys, demonstrating competitive cold-start performance with substantially lower training costs.
References:
[1] Wang et al. Learnable Item Tokenization for Generative Recommendation. CIKM, 2024.
[2] Liang et al. Rethinking Generative Recommender Tokenizer: RecSys-Native Encoding and Semantic Quantization Beyond LLMs. arXiv:2602.02338, 2026.
[3] Cheng et al. CapsID: Soft-Routed Variable-Length Semantic IDs for Generative Recommendation. arXiv:2605.05096, 2026.
[4] Xia et al. Unleash the Potential of Long Semantic IDs for Generative Recommendation. arXiv:2602.13573, 2026.
[5] Huang et al. Asymmetric Generative Recommendation via Kronecker Residual Bridge and Multi-Faceted Hierarchical Quantization. arXiv:2605.14512, 2026.
Thank you for your valuable feedback. We will revise the manuscript accordingly. We hope our responses have addressed your concerns and that you would consider revisiting your rating.
W1: Technical Novelty
We would like to clarify that Cuber's novelty lies beyond simply combining existing techniques. Our contributions address three key challenges in parallel generative recommendation.
First, RPQ-Tree extends RQ–PQ quantization to parallel generative recommendation. RVPQ [1] was originally developed for approximate nearest neighbor (ANN) search, which relies on quantized representations for distance-based retrieval. In contrast, generative recommendation requires predicting SID codes from user histories, where lower reconstruction error does not necessarily imply better recommendation performance. Our work demonstrates the effectiveness of combining RQ and PQ for single-forward SID prediction. Furthermore, unlike the fixed PQ-then-RQ structure of RVPQ, RPQ-Tree enables flexible recursive compositions, including different subspace partitions across residual stages.
Second, geometry-aware initialization provides a lightweight bridge between SID construction and recommendation learning. Conventional GR methods typically separate these stages, preventing the generator from directly exploiting the codeword geometry learned during quantization. While methods such as ETEGRec and DIGER address this gap through joint optimization, our approach transfers the learned geometric relationships into token embeddings through a simple, geometry-preserving initialization, avoiding additional joint-training complexity.
Third, progressive training exploits the SID hierarchy rather than merely applying conventional hard-negative mining. It progressively uses deeper tree levels to identify increasingly fine-grained confusions and incorporates them into both code-level discrimination and catalog-level ranking.
Together, these contributions form a unified framework in which SID construction, geometric initialization, and hierarchy-guided training are jointly designed to improve recommendation quality while retaining efficient parallel prediction.
W2: Effectiveness of Code Dependencies
Cuber exploits the structural dependencies introduced during SID construction without explicitly conditioning one prediction head on another. Although SID codes are predicted in parallel, the prediction heads share the same user-history representation and are jointly optimized through a shared backbone. We hypothesize that this shared representation helps the model learn to predict hierarchically structured SIDs without requiring explicit interactions between the heads.
During the initial design of our method, we also explored explicit dependency modeling through an MTP-inspired layer-wise predictor (Cuber-LP) on the two-level Cuber-A and Cuber-B variants. It first predicts higher-level code distributions and then uses their probability-weighted embeddings to condition lower-level predictions through an additional structured attention layer. As shown below, Cuber-LP does not yield consistent improvements across the three datasets, while increasing trainable parameters by approximately 12.8%.
Given these mixed results and the additional complexity, we retain RPG's parallel multi-head prediction architecture to preserve inference efficiency. Nevertheless, explicit dependency modeling remains a direction worth further investigation.
The ablation study in Table 3 already provides evidence for the contribution of hierarchical SID construction. Specifically, it compares OPQ-32 with RPQ-Tree before introducing geometry-aware initialization or progressive training, using the same co-occurrence-augmented item representations and parallel prediction architecture. RPQ-Tree consistently outperforms OPQ-32 across all three datasets, supporting the contribution of hierarchical dependencies introduced during SID construction to recommendation performance, independently of the other two proposed components.
References:
[1] Niu et al. Residual Vector Product Quantization for Approximate Nearest Neighbor Search. Expert Systems with Applications, 2023.