- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies task-adapted semantic IDs for sequential recommendation. Starting from fixed sentence-T5 metadata embeddings, it jointly trains an item tokenizer and a masked diffusion recommender, using a straight-through estimator to propagate recommendation gradients to the tokenizer. Its ELBO-motivated training objective combines denoising and catalog-retrieval losses with auxiliary reconstruction and vector-quantization losses.
The proposed beam-free inference first scores valid catalog items from a fully masked target, then rescores 100 candidates using all nonempty target-mask patterns. A validation-fitted ranker combines these neural scores with training-derived transition and popularity features. The final discrete pipeline also includes collision repair followed by denoiser finetuning with the tokenizer frozen. Continuous and hybrid latents are evaluated as representation ablations rather than components of the selected system. Experiments on three Amazon Reviews 2023 categories report higher NDCG than the compared baselines, and an ablation on Musical Instruments supports the benefits of joint tokenizer training.
Critical strengths
-
The paper addresses a meaningful mismatch between content-based tokenization and the downstream recommendation objective. Allowing recommendation gradients to update the tokenizer is well motivated, and Table 3 provides direct supporting evidence: on Musical Instruments, joint training improves NDCG@10 from 0.0290 to 0.0300 and reduces the fraction of catalog items with colliding IDs from 1.12% to 0.45%.
-
The catalog-supported proposal and denoising-based rescoring provide a useful alternative to sequential beam expansion. The proposal ensures that candidates correspond to catalog items, while the denoising score evaluates compatibility under multiple masking patterns. Table 1 reports the highest NDCG among the compared methods on all three datasets.
Critical weaknesses
-
The evidence isolating the contribution of joint representation learning remains incomplete. I appreciate that Table 3 includes a joint-versus-frozen comparison and Table 6 decomposes the ranking components. However, the controlled joint-training comparison is reported only on Musical Instruments, with a modest NDCG@10 gain and no uncertainty estimates. The final ranker also has a substantial effect: Table 6 increases NDCG@10 from 0.0327 with denoising scores alone to 0.0366 with the complete ranker. Since the external baselines in Table 1 are taken from Shi et al. (2025), the available evidence does not fully isolate tokenizer adaptation from differences in ranking and training stages. Moreover, closely related approaches to task-adapted tokenization, such as ETEGRec and DIGER, are discussed in Section 5 but absent from the empirical comparison. These gaps limit the evidence for the specific contribution beyond the complete system; the three Amazon categories also provide limited evidence across domains and catalog scales.
-
The assumptions behind Eq. (13) need clarification. The Bayes decoder in Eq. (10) gives , where . This denominator generally remains inside the marginalization over . The paper already uses a uniform item prior and hard IDs, and discusses uniqueness in Section 3.1. If the IDs are unique, is constant on valid catalog IDs and the ranking simplification is justified. However, residual collisions or soft assignments can make it depend on and change the ranking if omitted. Please clarify whether the final repaired catalog has unique IDs and how the proposal and denoising scores in Eqs. (14)-(15) relate to this derivation.
Questions for authors
-
Under what assumptions is the normalization term removable in Eq. (13), and does the final system satisfy those assumptions after collision repair? Please clarify how Eqs. (14)-(15) approximate the resulting item-ranking objective, including the treatment of any remaining collisions.
-
Can the joint-versus-frozen comparison be extended across the three datasets with matched training budgets, collision handling, decoding, and ranking, together with variability across runs? The most informative additional controls would be applying the same ranking procedure and validation budget to the frozen/LLaDA-Rec counterpart, and comparing against a directly relevant task-adapted tokenizer baseline such as ETEGRec or DIGER. Please also confirm that the imported baseline numbers use compatible splits, candidate catalogs, and evaluation filtering.
-
At which stage are the Table 3 results measured? Please report how recommendation quality and collision rate change after joint training, after repair, and after the 50-epoch frozen continuation, along with their training costs. Does the frozen baseline receive a comparable continuation budget? The final frozen stage does not contradict joint learning, but its contribution should be distinguished from that of tokenizer adaptation.
-
What are the measured latency, throughput, and peak memory of beam-free retrieval versus native beam decoding under matched hardware, batching, and recommendation quality?
3 comments
At which stage are the Table 3 results measured? Please report how recommendation quality and collision rate change after joint training, after repair, and after the 50-epoch frozen continuation, along with their training costs. Does the frozen baseline receive a comparable continuation budget? The final frozen stage does not contradict joint learning, but its contribution should be distinguished from that of tokenizer adaptation.
In Table 3, we report results from the final checkpoints, after the 50-epoch frozen continuation. Both models receive the same training compute, i.e., 150 epochs of denoiser training followed by 50 additional epochs with the tokenizer frozen.
The only difference is that for the joint model, the tokenizer and denoiser are jointly optimized during the 150-epoch stage, while for the frozen baseline, the tokenizer is pretrained using the reconstruction and quantization objectives. The tokenizer remains frozen throughout both denoiser-training stages.
Both rows use neural-only beam decoding without the reranker.
What are the measured latency, throughput, and peak memory of beam-free retrieval versus native beam decoding under matched hardware, batching, and recommendation quality?
We measured both methods on an NVIDIA A100 (40 GB). The times include generation, denoising-based scoring, and reranking.
| Dataset | Method | Latency (ms, batch 1) | Throughput (queries/s, batch 64) | Peak GPU memory (GiB, batch 64) |
|---|---|---|---|---|
| Industrial & Scientific | Native-beam | 140.4 | 7.63 | 10.74 |
| Industrial & Scientific | Beam-free | 108.0 | 9.78 | 1.90 |
| Musical Instruments | Native-beam | 153.2 | 6.87 | 10.73 |
| Musical Instruments | Beam-free | 117.8 | 8.81 | 1.89 |
| Video Games | Native-beam | 222.2 | 4.75 | 10.75 |
| Video Games | Beam-free | 172.6 | 5.99 | 1.91 |
We again thank the reviewer for their feedback.
Thank you for the additional experiments. Jointly training the tokenizer improves NDCG@10 on all three datasets with neural-only beam decoding. In the full pipeline, the differences relative to a frozen tokenizer are small and mixed +0.0005, −0.0002, and −0.0004. Since joint training is a central contribution and its benefit in the full system remains unclear, I am keeping my rating at 2.
We thank the reviewer for their feedback. In the following, we address their points.
We agree that the frozen-tokenizer ablation alone does not establish an advantage over existing task-adapted tokenizers. Therefore, we evaluated DIGER on all three datasets using the same splits, training catalog, and T5 features.
In the following table, we report NDCG@10 of the test split:
As we observe, our method achieves higher NDCG@10 in these comparisons.
We acknowledge that joint tokenization and recommendation have been explored previously. However, our contribution focuses on combining semantic-ID learning with masked diffusion within a variational framework and our beam-free retrieval.
In general, the denominator rϕ(z) depends on z and cannot be omitted. With a uniform item prior and deterministic IDs, the correct ranking is:
p(y∣x)∝m(z(y))pθ(z(y)∣x),
where m(z) counts items sharing the ID z. However, the approximation is correct when there are no collisions, which is the case for us as our final collision rate is close to 0.
We will clarify this assumption in the methodology section.
We follow the reviewer's suggestion and extend the joint-versus-frozen comparison to all three datasets, using both neural-only beam decoding and complete inference. Both training procedures include collision repair.
The following table reports NDCG@10:
As we observe, joint training improves neural-only performance across all three datasets. However, when including the ranker, the performance becomes numerically close.
We have confirmed that the splits are equal. While the Llada-Rec implementation used all items for their catalog, we use only the training catalog and count targets outside that catalog as misses.