acceptodds
Under review as a conference paper at ICLR 2027

SPRINT: Single-Step Generative Recommendation via Average Probability Velocity

Abstract

Semantic ID (SID) based generative recommendation represents each item as a sequence of multiple discrete tokens, and recommends by generating the SID of the item a user would like to interact with. Both dominant paradigms in this domain generally pay for generation token by token: autoregressive models decode the tokens left-to-right, while non-autoregressive models decode in parallel yet still need multiple rounds of refinement to stay competitive. Therefore, both generally spend multiple forward passes per item, a cost that is prohibitive in latency-sensitive recommender systems. We ask whether an item can be generated in a single forward pass, and answer it through a new perspective which we call average probability velocity. We view SID generation as a flow of token generation probabilities and characterize it by its average velocity over the whole generation process. We prove that this average velocity is fully determined by the average generation probability of each token. Therefore, we directly parameterize and learn the probabilities of all tokens in a single forward pass with a bidirectional Transformer. As these probabilities are generated independently across positions and the coherence among tokens is lost, we further design a dual-level flow contrastive objective to restore the coherence among an item's tokens. It contrasts the target SID against negative SIDs at both the token and SID levels. The token level ranks the generation probabilities of the target tokens above those of negative SIDs, while the SID level scores the tokens of each SID as a whole item for capturing token coherence of each item. Extensive experiments on 8 real-world datasets show that our model not only generates recommendations far more efficiently ( speedup over the second-fastest AR/NAR method) but also attains superior recommendation accuracy (% average improvement over the second-best).

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Generative Recommendation, Non-autoregressive generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_meP8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPRINT, which applies MeanFlow’s average-velocity perspective to discrete semantic ID (SID) generation. It derives average token posteriors from the average probability velocity and uses a single Transformer forward pass to predict a complete SID. The model combines token-level contrastive learning with SID-level contrastive matching. The paper evaluates recommendation accuracy and inference efficiency on eight datasets.

Critical strengths
  1. The problem is practically relevant: autoregressive decoding and iterative denoising increase inference latency in generative recommendation.

  2. The two-level objective addresses position-wise prediction and whole-SID matching; the ablations show that both components contribute.

  3. The experiments cover multiple datasets and report recommendation accuracy, inference efficiency, and component ablations.

Critical weaknesses
  1. The connection between the central theory and the optimized objective is unclear. The average-probability-velocity derivation leads to a cross-entropy objective over average token posteriors. The final objective replaces this with token-level candidate ranking and adds an SID-level matching term that is not derived from APV. Inference also does not explicitly compute or integrate a velocity. The paper therefore does not establish what APV contributes beyond direct one-step SID prediction. A matched baseline using the same architecture and training objective, but trained as a direct one-step SID predictor, would help isolate this contribution.

  2. The sources of the performance gains are not adequately separated. The ablations show that the SID-level branch contributes substantially, but the paper does not add this matching branch to baselines such as RPG or OneGR. Moreover, w/o Contrast removes both negative-sample contrast and SID-level matching, so it cannot measure their individual contributions.

  3. The scope of the inference-efficiency claims needs clarification. A single Transformer forward pass does not make catalog search independent of the number of items. The released code scores every catalog item before returning the top K, so computation and cached-item memory grow with catalog size. Given that methods such as RPG use graph-constrained search, the paper should compare graph- or ANN-based candidate search, or clearly scope its efficiency claims to reranking a pre-retrieved candidate set.

  4. The released code does not fully match the reported experimental settings. The paper reports \rho=0.75 for Scientific and Instruments, whereas the code fixes the training proportion at 1.0. The code also uses an SID-level softmax temperature that is not reported in the paper’s equations or hyperparameter table. In addition, the paper reports results averaged over five random seeds, but the released configuration and run instructions do not provide a corresponding multi-seed procedure.

  5. The negative-sampling implementation may not match the paper’s definition. The paper describes negatives as items the user has not interacted with. The code samples with replacement from the global catalog and excludes only the current target SID; it does not filter the user’s history. Thus, previously interacted items may be used as negatives.

Questions for authors

Please refer to Weaknesses.

9 comments
AuthorsOpenReview ↗

Responses to Reviewer meP8 (Part 1)

Thank you for your thorough review and for carefully examining our released code. We appreciate your recognition of the practical relevance of the problem, the complementary roles of the two objective levels, and the breadth of our experiments. Below are our detailed responses. We would be glad to discuss further.

Weakness 1: Connection between the central theory and the optimized objective: "The final objective replaces this with token-level candidate ranking and adds an SID-level matching term that is not derived from APV. Inference also does not explicitly compute or integrate a velocity."

Thank you for this comment. Our objective is derived from the average probability velocity (APV), and we have added two formal results in the updated version to make this explicit (refer to Appendix C.5 and C.6 in Page 23-25 in our updated version).

(1). The token-level objective is APV matching.

Eq. (8) aligns the marginal and conditional APVs (refe to Appendix C.5). Subtracting the marginal APV (Eq. 6) from the conditional one (Eq. 5), the current-state terms cancel and the shared scalar factors out:

vˉr,tl(xl,z∣x1)−vˉr,tl(xl,z)=λr,t[δx1(xl)−pˉ1∣r,t(xl∣z)].\bar v^l_{r,t}(x^l,z\mid x_1)-\bar v^l_{r,t}(x^l,z)=\lambda_{r,t}\big[\delta_{x_1}(x^l)-\bar p_{1|r,t}(x^l\mid z)\big].

The velocity residual thus reduces to a posterior residual. Any consistent velocity-matching loss on this residual and the per-token cross-entropy in Eq. (8) therefore share the same minimizer. Eq. (8) is a valid surrogate for APV matching, not a replacement of it.

Eq. (9) strictly extends Eq. (8) . Let Φ(c)=∑llog⁡pˉ1∣r,t(cl∣xr,H,θ)\Phi(c)=\sum_l\log\bar p_{1|r,t}(c^l\mid x_r,\mathcal{H},\theta). Then

∇θLCL=∇θLCE+∑c∈Sw(c)∇θΦ(c),w(c)=exp⁡(Φ(c))∑c′∈Sexp⁡(Φ(c′)).\nabla_\theta\mathcal{L}_{\mathrm{CL}}=\nabla_\theta\mathcal{L}_{\mathrm{CE}}+\sum_{c\in\mathcal{S}}w(c)\nabla_\theta\Phi(c),\qquad w(c)=\frac{\exp(\Phi(c))}{\sum_{c'\in\mathcal{S}}\exp(\Phi(c'))}.

The contrastive gradient contains the full APV-matching gradient, plus a term that suppresses competing SIDs. The contrastive form therefore builds on the APV objective rather than replacing it.

(2). The SID-level term is motivated by the structure of the one-step transition.

The exact one-step transition from the all-mask state is the joint posterior over the whole SID, while a single parallel pass recovers only its per-position marginals. The factorized score therefore cannot express how the tokens of an SID constrain one another. The SID-level term scores the LL tokens jointly as one item and supplies exactly this missing dependence. We have explicitly explain this in our manuscript (refer to Sec. 3.4 Lines 310-315 in Page 6)

(3). Inference is the closed-form integration of the APV.

Because the APV has a fixed direction and a closed-form rate (our Proposition 2 in Appendix C.3), it can be integrated exactly. The per-position transition is Pt∣r=δz+(1−e−(t−r)λr,t)(pˉ1∣r,t−δz)P_{t|r}=\delta_z+(1-e^{-(t-r)\lambda_{r,t}})(\bar p_{1|r,t}-\delta_z), which reduces to pˉ1∣0,1\bar p_{1|0,1} on [0,1][0,1]. Reading pˉ1∣0,1\bar p_{1|0,1} in Eq. (11) is therefore the integrated APV, and no numerical integration is needed.

AuthorsOpenReview ↗

Responses to Reviewer meP8 (Part 2)

Weakness 2: Separating the sources of the gains: "the paper does not add this matching branch to baselines such as RPG or OneGR. Moreover, w/o Contrast removes both negative-sample contrast and SID-level matching."

Thank you for this suggestion. We add both experiments.

(1). RPG has better performance with our SID-level flow objective.

We have added our SID-level branch to RPG, Table 1 shows the results. From the results, we can have the following finding:

Key Finding: RPG can achieve better performance with our SID-level flow objective.

Table 1. Performance comparision between RPG and RPG + SID-level flow objective.

MethodSportsBeautyToysScientificInstrumentsGames
R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10
RPG0.04630.02630.08090.04640.08690.04900.03950.02180.05450.03000.08530.0485
RPG + SID-level flow objective0.05030.02790.08650.04900.09180.04940.04540.02450.06040.03260.09200.0504
Improv.8.69%6.25%6.91%5.62%5.59%0.88%14.94%12.39%10.83%8.67%7.85%3.92%

(2). Contrastive formulation is necessary and contribute positively to our training.

To address your concern, I have design three variants :

  • V1: replacing the token-level contrastive objective with ordinary per-token cross-entropy while keeping others unchanged.
  • V2: replacing the SID-level contrastive objective with ordinary cross-entropy while keeping others unchanged.
  • V3: replacing both token-level and SID-level contrastive objective with cross-entropy.

Table 2 reports the performance of our method and three variants. We have the following finding.

Key Finding: Removing either side contrastive objective harms model performance to a great extent.

Table 2. Performance comparison between different variants in terms of contrastive formulation.

MethodSportsBeautyToysScientificInstrumentsGames
R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10
SPRINT (ours)0.06210.03470.09870.05520.09980.05940.06560.03690.07130.03860.10010.0534
V10.04660.02470.08210.04630.09000.05470.04820.02800.05150.02800.07450.0408
V20.04420.02340.08310.04600.08860.05440.04320.02480.04740.02590.06960.0374
V30.02830.01550.06450.03810.07560.04770.03070.01870.03240.01880.05370.0310
AuthorsOpenReview ↗

Responses to Reviewer meP8 (Part 3)

Weakness 3: Scope of the inference-efficiency claims: "A single Transformer forward pass does not make catalog search independent of the number of items ... compare graph- or ANN-based candidate search, or clearly scope its efficiency claims."

Thank you for this comment. We have provided detailed complexity analysis of graph-constrained search of RPG and ANN-based candidate search of our method SPRINT below.

RPG. RPG's search and scoring consist of two parts. The first is the construction of its decoding graph. For each of the NN valid SIDs, RPG retrieves kk nearest neighbours, and each similarity costs O(m)O(m), where mm is the SID length. This gives O(N2m)O(N^2m) time and O(Nk)O(Nk) memory. The graph must also be maintained as the catalog changes. Each new item requires O(Nm)O(Nm) time to find its neighbours, and the adjacency lists of those neighbours must then be updated. The second part is online decoding. Starting from a beam of bb SIDs, RPG runs qq iterations. In each iteration, it expands the kk neighbours of every beam SID, scores each candidate with mm additions, and keeps the top bb. This costs O(qbk(m+log⁡b))O\big(qbk(m+\log b)\big) per user. The overall complexity is therefore O(N2m+qbk(m+log⁡b))O\big(N^2m+qbk(m+\log b)\big). The qq iterations must run sequentially, and each one involves irregular memory access over the graph.

SPRINT. SPRINT requires no decoding graph. After one forward pass, it scores all NN valid SIDs directly. The token-level term sums LL table lookups per SID, which costs O(NL)O(NL). The SID-level term is a single matrix product between u\mathbf{u} and the precomputed {vc}\{\mathbf{v}_c\}, which costs O(Nd)O(Nd). The overall online complexity is therefore O(N(L+d))O\big(N(L+d)\big). All of these operations are dense and fully parallel, and they finish in one round. The SID representations {vc}\{\mathbf{v}_c\} are computed once offline in O(NLd2)O(NLd^2). A new item only needs one MLP pass to compute its vc\mathbf{v}_c, at a cost of O(Ld2)O(Ld^2), and its representation is then appended, with no graph to rebuild or update.

Comparison.

(1) Offline preparation.

RPG builds its decoding graph in O(N2m)O(N^2m), while SPRINT precomputes {vc}\{\mathbf{v}_c\} in O(NLd2)O(NLd^2). RPG is therefore more expensive whenever N>Ld2/mN>Ld^2/m, which holds for typical catalog sizes, and the gap grows linearly with NN.

(2) Catalog updates.

For each new item, RPG needs O(Nm)O(Nm) to find its neighbours, while SPRINT needs O(Ld2)O(Ld^2) to compute its vc\mathbf{v}_c. RPG is therefore more expensive whenever N>Ld2/mN>Ld^2/m, and its cost keeps growing with the catalog, whereas SPRINT's cost stays constant.

(3) Online scoring per user.

RPG’s online decoding costs O(qbk(m+log⁡b))O\big(qbk(m+\log b)\big), independent of NN, but it relies on a decoding graph that costs O(N2m)O(N^2m) to build. SPRINT needs no graph. It scores all NN SIDs in O(N(L+d))O\big(N(L+d)\big) with a single, fully parallel matrix product, whereas RPG requires qq sequential rounds of irregular graph access. This is why SPRINT is substantially faster in practice (Table 3). Moreover, as stated in our manuscript (refer to Appendix F), SPRINT is more suitalbe in the ranking stage, where a candidate set of size Nc≪∣I∣N_c\ll|\mathcal{I}| has already been retrieved. There, SPRINT costs only O(Nc(L+d))O\big(N_c(L+d)\big), whereas RPG would have to rebuild its graph for every candidate set at O(Nc2m)O(N_c^2m). SPRINT therefore has a lower complexity than RPG in the ranking stage.

Reviewer_wfDnOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPRINT for single-step generative recommendation. The method predicts all four SID tokens in parallel and combines token-level and SID-level contrastive objectives, motivated by an average probability velocity formulation. Experiments on eight datasets show improvements over the compared methods under full-catalog evaluation. The authors also study the effects of the tokenizer, contrastive objectives, and training mask ratio.

Critical strengths
  1. This paper provides a compact recommendation model that predicts four SID tokens in a single forward pass without iterative refinement.

  2. The contrastive objectives combine token prediction with matching at the item level. The ablation studies show the contributions of both scoring branches.

  3. The experiments cover eight datasets with full-catalog evaluation and report results over multiple seeds with significance tests.

Critical weaknesses
  1. RPG [1] already predicts all SID tokens in parallel in a single forward pass. The basic decoding strategy is therefore similar, and the methodological advance over existing parallel generation methods needs to be made clearer.

  2. The ablations show a large contribution from the SID-level matching branch, but do not isolate the benefit of the average probability velocity formulation. It is unclear what this formulation adds to a conventional parallel predictor trained with the same contrastive objectives.

  3. Inference still requires scoring all valid SIDs and computing similarities to their precomputed embeddings.

  4. This paper does not report complete-SID collision rates or explain how collisions are handled. If multiple items share the same SID, both scoring branches assign them identical scores. Please clarify how these SIDs are mapped back to individual items and how ties and hits are handled during evaluation.

[1] Hou et al. 2025. Generating Long Semantic IDs in Parallel for Recommendation. KDD 2025.

Questions for authors
  1. Can the authors isolate the benefit of the average probability velocity formulation using a parallel predictor with the same architecture, SID length, contrastive objectives, and inference procedure?

  2. What are the complete-SID collision rates for each dataset, and how are collisions and tied scores handled during evaluation?

9 comments
AuthorsOpenReview ↗

Responses to Reviewer wfDn (Part 1)

Thank you for your insightful review and for recognizing the compactness of our single-pass model, the complementary roles of the two scoring branches, and our evaluation protocol. Below are our detailed responses. We would be glad to discuss further.

Weakness 1: Methodological advance over RPG: "RPG already predicts all SID tokens in parallel in a single forward pass. The basic decoding strategy is therefore similar, and the methodological advance over existing parallel generation methods needs to be made clearer."

Thank you for this comment. SPRINT advances RPG in three key aspects.

(1). Short SIDs: 16-64 tokens per item (RPG) vs. 4 tokens per item (ours).

RPG relies on long SIDs of 16 to 64 tokens for expressiveness. This enlarges the embedding table by 4×4\times to 16×16\times (we have explicitly explain this in our manuscript, please refer to Lines 78-81 in our updated version) and requires 8.53M to 23.69M parameters. Ours method SPRINT uses compact 4-token SIDs and requires only 5.59M parameters. A compact token embedding table is important in industrial deployment. The embedding table accounts for a large share of the parameters of an SID-based GR model, and its size grows with the SID length. Keeping it small reduces memory usage and makes the model easier to deploy.

Moreover, when RPG represents each item with the same number of tokens as SPRINT (4 tokens), its performance drops significantly (R@10 drops from 0.0545 to 0.0183 in Instruments dataset), as shown in Table 1 below. This further confirms the advantage of SPRINT over RPG.

Key finding: With the same 4-token SIDs as SPRINT, RPG degrades sharply, so RPG depends on long SIDs, while SPRINT stays effective with compact ones.

Table 1: Comparison of RPG and SPRINT with different numbers of tokens per item (R@10 and N@10).

MethodSportsBeautyToysScientificInstrumentsGames
R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10
RPG (16-64 tokens)0.04630.02630.08090.04640.08690.04900.03950.02180.05450.03000.08530.0485
RPG (4 tokens)0.02590.01380.04890.02770.05670.03020.01750.00900.01830.01030.03240.0186
SPRINT (4 tokens)0.06210.03470.09870.05520.09980.05940.06560.03690.07130.03860.10010.0534

(2). Efficient decoding: Decoding graph over valid SIDs (RPG) vs. No decoding graph required (ours).

RPG decodes by propagating a beam over a decoding graph built on all valid SIDs. This graph must be constructed in advance and updated whenever new items arrive. SPRINT needs no such graph. It directly scores all valid SIDs after one forward pass, so a new item only needs its SID and its SID representation to be added.

(3). Effective learning objective: Token cohenrence is lost in learning (RPG) vs. Token coherence is explicitly learned through SID-level flow objective (ours).

RPG trains with a multi-token prediction loss, i.e., a sum of position-wise cross-entropy losses. The coherence among the tokens of an SID is therefore lost during training. SPRINT includes an SID-level flow that scores the tokens of each SID jointly, which captures this token coherence during training.

We have also provided a detail discussion in our updated version (refer to Lines 346-368 in Page 7 in our updated version).

AuthorsOpenReview ↗

Responses to Reviewer wfDn (Part 2)

Weakness 2 & Question 1: Isolating the benefit of the average probability velocity formulation: "The ablations show a large contribution from the SID-level matching branch, but do not isolate the benefit of the average probability velocity formulation. It is unclear what this formulation adds to a conventional parallel predictor trained with the same contrastive objectives. Can the authors isolate the benefit of the average probability velocity formulation using a parallel predictor with the same architecture, SID length, contrastive objectives, and inference procedure?"

Thank you for this question. To exame whether our average probability velocity formulation can benefit existing parallel GR model, we have conducted an exepriment by adding our average probability velocity formulation (token-level flow) to a representative and SOTA parallel GR: RPG. From the results shown in Table 2 below, we have the following finding.

Key Finding: Our average probability velocity formulation consistently improves existing parallel GR models, boosting RPG on all six datasets.

Table 2. Performance comparision between RPG and RPG+Average Probability Velocity formulation.

MethodSportsBeautyToysScientificInstrumentsGames
R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10
RPG0.04630.02630.08090.04640.08690.04900.03950.02180.05450.03000.08530.0485
RPG + Average Probability Velocity0.04900.02740.08410.04770.08900.04930.04220.02310.05790.03150.08920.0492
Improv.5.87%4.17%4.01%2.81%2.47%0.66%6.84%5.96%6.24%5.00%4.57%1.44%

Weakness 3: Inference still scores all valid SIDs: "Inference still requires scoring all valid SIDs and computing similarities to their precomputed embeddings."

Thank you for this comment. We respectfully clarify that this concerns the application scope of SPRINT rather than a weakness, and we have stated this scope explicitly in our manuscript.

(1). This is about the application scope, not a weakness.

In the Limitations section (Appendix F), we explicitly state that, because SPRINT scores all valid SIDs, it is best suited to the ranking stage. There, a candidate set has already been retrieved, and the goal is to produce a fine-grained recommendation list rather than to retrieve directly from billions of items. Within this scope, the number of valid SIDs is bounded by the candidate set. Scoring them takes a single, fully parallel matrix product, which accounts for only about a very small propotion of SPRINT's inference time (refer to Table 4 in our response to Reviewer ZMtL, we also post it in below Table 3).

Table 3: Breakdown of SPRINT's inference time (ms/sample; mean (min to max) over five runs).

Dataset#ItemsNetwork passScoring + top-KKTotal
Toys11,9242.972 (2.926 to 2.987)0.576 (0.571 to 0.583)3.548
Beauty12,1012.991 (2.908 to 3.153)0.581 (0.570 to 0.583)3.572
Sports18,3572.959 (2.938 to 2.989)0.590 (0.583 to 0.593)3.549
Instruments24,5872.959 (2.931 to 2.991)0.607 (0.607 to 0.614)3.567
Games25,6122.963 (2.937 to 2.982)0.605 (0.600 to 0.608)3.568
Scientific25,8482.972 (2.968 to 3.011)0.606 (0.603 to 0.613)3.578

(2). Many existing GR methods share this paradigm.

Scoring or searching over the set of valid SIDs is common in GR [1-4]. This scope is a property shared by the paradigm, not a specific weakness of SPRINT.

[1] Lin X, Shi H, Wang W, Feng F, Wang Q, Ng SK, Chua TS. Order-agnostic identifier for large language model-based generative recommendation. In SIGIR 2025 (pp. 1923-1933).

[2] Wang P, Zhang Y, Zhang R, Guo J, Cheng X. One-Pass Decoding for Generative Recommendation with WFST-Constrained A* Search. In SIGIR 2026 (pp. 4273-4278).

[3] Qu H, Fan W, Zhao Z, Li Q. Tokenrec: Learning to tokenize id for llm-based generative recommendations. IEEE Transactions on Knowledge and Data Engineering. 2025 Aug 19.

[4] Xia M, Ma G, Zhou Z, Huang D. Unleash the potential of long semantic ids for generative recommendation. arXiv preprint arXiv:2602.13573. 2026 Feb 14.

AuthorsOpenReview ↗

Responses to Reviewer wfDn (Part 3)

Weakness 4 & Question 2: SID collisions and tie handling: "What are the complete-SID collision rates for each dataset, and how are collisions and tied scores handled during evaluation?"

Thank you for this careful question. We aim to address your concerns through following four perspectives.

(1). SID construction strictly follows prior work.

We follow the practice of RPG, DiffGRM, and OneGR and use OPQ as the tokenizer to construct SIDs. Like these works, as well as LLaDA-Rec and many other GR models, we do not apply special handling to SID collisions. In real-world recommender systems, there are supporting solutions for this, such as policy-based re-ranking in later stages, which determine the final item to display among those sharing an SID. Collision handling is therefore not the research focus of this work, nor of most GR studies.

Table 4 reports the collision statistics. The same-category rate is the proportion of items sharing the same SID that belong to the same fine-grained category. The strongest baseline methods DiffGRM and OneGR share the same collision rate.

Table 4: SID collision and category statistics.

DatasetsSportsBeautyToysScientificInstrumentsGamesArtsYelp
Collision rate17.73%14.25%14.29%23.36%14.79%18.23%18.28%18.94%
Rate of same category96.50%94.00%91.60%98.40%97.20%82.40%96.00%99.90%

Most collisions (82.40% to 99.90%) occur among items from the same category. Since OPQ quantizes semantic embeddings, colliding items are typically closely related, and collisions mainly group similar items rather than unrelated ones.

(2). The evaluation protocol is consistent across all GR baselines.

Evaluation is conducted directly in the SID space, i.e., a prediction is counted as correct when the generated SID matches the SID of the ground-truth item. This is a common protocol in this line of research. Since scores are computed per SID, items sharing an SID form a single prediction and do not introduce ties among distinct predictions. All generative baselines in our comparison are evaluated under the same protocol, so the comparison among them is consistent and fair.

(3). Collisions among similar items are a characteristic of SIDs.

SID collisions are not necessarily harmful. Some methods regard them as beneficial, because they cluster semantically very similar items together. For example, QuaSID [5] argues that not all collisions are equally harmful, and treats collisions among related items differently from those among unrelated ones. Grouping similar items under shared codes is a key characteristic of SIDs, which allows knowledge to be shared among related items. Table 4 is consistent with this view, since most collisions in our datasets occur within the same category.

[5] Hu Z, Chen Y, Pan Y, Yuan X, Yin Y, Wang D, Xia B, Luo Z, Wang H, Ni S, Liang D. Stop treating collisions equally: Qualification-aware semantic id learning for recommendation at industrial scale. In KDD 2026 (pp. 7401-7411).

(4). Our method SPRINT remains the best even under 0% collision setting.

To further rule out any effect of collisions, we evaluate the generative methods with collision-free SIDs (0% collision) on Scientific, Instruments, and Games (Table 5). Specifically, we append one extra disambiguation token to every SID. Items that share an SID receive distinct values of this token, so that each SID maps to exactly one item. Items without collisions also receive this extra token, so that all items keep the same SID length.

Table 5: Results with collision-free SIDs (0% collision).

DatasetScientificInstrumentsGames
MetricR@5N@5R@10N@10R@5N@5R@10N@10R@5N@5R@10N@10
TIGER0.02820.01830.04460.02360.03590.02330.05660.03000.05290.03480.08230.0442
COBRA0.02720.01790.04240.02280.03440.02310.05090.02850.04770.03070.07500.0395
RPG0.02570.01740.03950.02180.03620.02410.05450.03000.05790.03970.08530.0485
DiffGRM0.02610.01660.04350.02230.03210.02320.05440.02890.05120.03420.08170.0440
OneGR0.02410.01430.04180.02020.03090.02120.04540.02650.04670.03110.06920.0394
SPRINT0.03020.02080.04830.02620.04080.02620.06100.03310.05850.04010.08970.0491

Key Finding. Our method SPRINT still achieves the best results compared with baselines under 0% collision setting. This confirms that its advantage does not come from SID collisions.

Reviewer_ZMtLOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

SPRINT aims to reduce the decoding cost of semantic-ID recommendation. It represents each item with four OPQ tokens and predicts them in parallel using a lightweight bidirectional Transformer. Training combines token-level and SID-level contrastive losses, while inference combines token probabilities with SID embedding similarity scores to rank valid items. The method is motivated by an average probability velocity formulation and evaluated on eight datasets.

Critical strengths
  1. Reducing the number of neural decoding passes while keeping semantic IDs short could be useful for recommendation systems with tight inference budgets.
  2. The results are consistently positive across the eight datasets in Tables 2, 8, and 12. Full-catalog evaluation, five random seeds, and significance tests provide reasonable support for the reported accuracy improvements.
Critical weaknesses
  1. The connection between the theory and the proposed algorithm is unclear. For the all-mask source and the masking path in Eq. (2), the target posterior conditioned on a fixed partially revealed state (z) is time-independent, as also established in Proposition 3.1 and Appendix E.8 of Gat et al., Discrete Flow Matching. The weighted average in Eq. (7) therefore reduces to the ordinary conditional denoising posterior wherever the integral is well-defined. This makes the claimed distinction between learning an interval-averaged posterior and learning a posterior at a particular time unclear.There is also a gap between averaging the generator at a fixed state and obtaining a finite-time, one-step generation rule. The paper acknowledges that the coefficient diverges at the terminal time and explains that it is not used in the implementation. This addresses a numerical concern but does not establish how the derivation leads to the final predictor.
  2. The ablations do not isolate the source of the gains. Table 7 uses (\rho=1) for six of the eight datasets, meaning that the best-performing models on those datasets are trained entirely from the all-mask state. Since RPG already uses OPQ and parallel token prediction, the contribution of the average-velocity formulation beyond the prediction architecture and the added SID-level objective remains unclear.Table 4 suggests that the SID branch accounts for a substantial part of the improvement. On Scientific, R@5 is 0.0444 for the full model, 0.0428 without the token branch, and 0.0313 without the SID branch. However, because the SID branch is used both during training and in the final ranking score, these results do not distinguish improvements in token representations or coherence from gains due to the additional ranking component. The current “w/o contrast” ablation removes more than one component and therefore cannot isolate the effect of the token-level contrastive objective.
  3. The efficiency claims need a clearer comparison protocol. The 8.39–10.04× speedup in the abstract is measured against RPG in Table 3. Table 10 includes a faster AR baseline, SID-MLP, against which the speedup is approximately 1.66–2.23×. The phrase “second-fastest AR/NAR method” therefore leaves the comparison set ambiguous.In addition, Appendix D.3 states that efficiency comparisons use standardized configurations, whereas accuracy comparisons use separately tuned configurations. Consequently, the reported results do not establish the accuracy–efficiency tradeoff of the same model configurations. Eq. (11) also requires scoring valid SIDs, so candidate scoring and top-(K) selection costs matter in addition to the single network pass. Although Appendix F acknowledges the dependence on catalog size, the timing boundaries and the inclusion of these costs need to be explicit.
Questions for authors
  1. For the all-mask source and masking path in Eq. (2), what does the posterior averaging in Eq. (7) add to the learning problem, given that the target posterior conditioned on a fixed partially revealed state is time-independent?
  2. How does averaging the generator at a fixed state lead to the implemented finite-time, one-step predictor, particularly when the divergent terminal coefficient is omitted?
  3. Given that six datasets use (\rho=1), what contribution does the average-velocity formulation make beyond OPQ-based parallel prediction and the SID-level objective?
  4. Could you provide a control that keeps the SID branch and inference procedure unchanged while replacing the token-level contrastive objective with ordinary per-token cross-entropy? This would help isolate the contribution of the token objective from that of SID-level matching and ranking.
  5. What comparison set does “second-fastest AR/NAR method” refer to, and how should the abstract’s speedup claim be interpreted in light of the faster SID-MLP baseline in Table 10?
7 comments
AuthorsOpenReview ↗

Responses to Reviewer ZMtL (Part 1)

Thank you for your detailed review. We appreciate your recognition of the practical value of our proposed model. Our detailed responses are below. We would be glad to discuss further.

Weakness 1 & Questions 1, 2: The distinction between learning an interval-averaged posterior and learning a posterior at a particular time.

Thank you for this insightful comment. The time invariance you point out is fully consistent with our derivation. In fact, it is the property that makes exact single-step generation possible. We explain this in two steps.

(1). The distinction (learning an interval-averaged posterior vs. learning a posterior at a particular time) is between two learning problems, not between two values at a fixed state (W1 & Q1).

We agree that, under our masking path, the posterior at a fixed state is τ\tau-invariant. Our claim is not that the two posteriors take different values at the same zz. The distinction lies in what the network represents and how it is trained.

  • What is represented. A posterior at a particular time parameterizes the local generator vτv_\tau (Eq. 3). It is valid only at the current state and must be re-evaluated as the state evolves. This needs one network forward per step, at the evolving states, to reach t=1t=1. In GR, this means LL network passes, where LL is the number of tokens per item. The interval-averaged posterior (ours) parameterizes the transition over the whole interval from the starting state. This only needs one network forward at the starting state.
  • How it is trained. The former is fitted on intervals that end at t=1t=1, with the start anchored at the all-mask or partial-mask state (refer to Sec. 3.3, "Time sampling" in Lines 281-290 in Page 6 in our updated version). The latter is fitted on states z∼pτz\sim p_\tau with τ∼U[0,1]\tau\sim\mathcal{U}[0,1].

The distinction has a large empirical effect.

  • Finding 1: As shown in Table 1, learning interval-averaged posterior (ours) consistently outperforms learning a posterior at a particular time in terms of recommendation accuracy.
  • Finding 2: Besides, as shown in Table 2, learning interval-averaged posterior (our) converges much faster than learning a posterior at a particular time.

Table 1: Performance comparision between learning a posterior at a particular time and Interval-averaged posterior (ours).

SportsBeautyToysScientificInstrumentsGames
R@5N@5R@5N@5R@5N@5R@5N@5R@5N@5R@5N@5
Posterior at a particular time0.03720.02450.06010.03990.06800.04820.04260.02910.04550.02970.06020.0391
Interval-averaged posterior (ours)0.04120.02700.06750.04520.07090.05020.04450.03020.04620.03060.06370.0417

Table 2: Covergence speed comparision ( the epoch to achieve the best performance) between learning a posterior at a particular time and Interval-averaged posterior (ours).

SportsBeautyToysScientificInstrumentsGames
Posterior at a particular time728112210867106
Interval-averaged posterior (ours)323137353637

(2). From the average velocity to the finite-time one-step predictor (W1 & Q2).

Integrating the average velocity exactly yields a valid finite-time transition, which reduces to the one-step predictor.

Explanation: The coefficient λr,t\lambda_{r,t} diverges at t=1t=1. This is the standard behaviour of a CTMC rate whose transition must complete by the terminal time [1]. The finite-time rule is therefore obtained by exact integration rather than by an Euler step, and the closed-form λr,t\lambda_{r,t} is what makes this integration possible. Given x1x_1, a masked position stays masked until tt with probability (1−κt)/(1−κr)=e−(t−r)λr,t(1-\kappa_t)/(1-\kappa_r)=e^{-(t-r)\lambda_{r,t}}. Marginalizing over the posterior gives the per-position transition over [r,t][r,t]:

Pt∣r(xl∣z)=δz(xl)+(1−e−(t−r)λr,t)[pˉ1∣r,t(xl∣z)−δz(xl)].P_{t|r}(x^l|z)=\delta_z(x^l)+\big(1-e^{-(t-r)\lambda_{r,t}}\big)\big[\bar p_{1|r,t}(x^l|z)-\delta_z(x^l)\big].

The average velocity thus determines both the direction pˉ−δz\bar p-\delta_z and the step size 1−e−(t−r)λr,t∈[0,1]1-e^{-(t-r)\lambda_{r,t}}\in[0,1]. Hence Pt∣rP_{t|r} is a valid distribution for every interval. As t→1t\to1, the divergent λr,t\lambda_{r,t} drives e−(t−r)λr,t→0e^{-(t-r)\lambda_{r,t}}\to0, and the rule becomes P1∣r(xl∣z)=pˉ1∣r,1(xl∣z)P_{1|r}(x^l|z)=\bar p_{1|r,1}(x^l|z). No coefficient is discarded. λr,t\lambda_{r,t} enters the finite-time rule only through this exponential, which equals one at the terminal time. The one-step predictor in Eq. (11) is therefore exactly the transition implied by the average velocity over time interval [0,1][0,1].

[1] Campbell A, Yim J, Barzilay R, Rainforth T, Jaakkola T. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design. In ICML 2024. (pp. 5453-5512).

AuthorsOpenReview ↗

Responses to Reviewer ZMtL (Part 2)

Weakness 2 & Question 3: Contribution of the formulation given ρ=1\rho=1 on six datasets: "Since RPG already uses OPQ and parallel token prediction, the contribution of the average-velocity formulation beyond the prediction architecture and the added SID-level objective remains unclear."

Thank you for this question. We summarize the contributions of the average-velocity formulation below.

Contributions of the average-velocity formulation.

  • (1) Theoretical foundation for exact single-step generation. The formulation shows that one network evaluation is the exact finite-time transition over time interval [0,1][0,1]. The fixed transport direction is exactly what makes a single interval-spanning step well-defined (the proof is already provided in Appendix C.3 in Page 22 in our updated version). This justifies the rationale of removing both time conditioning and iterative refinement.
  • (2) Our formulation is flexible, of which ρ=1\rho=1 is one member of our interval family. The formulation defines a family of training intervals [r,1][r,1] indexed by ρ\rho . ρ=1\rho=1 trains the exact map used at inference. With ρ<1\rho<1, the remaining fraction supplies conditional supervision in which the model predicts the remaining positions of an item given those already committed (refer to Lines 281-290 in Page 6 in our updated version). The best ρ\rho is data-dependent: 0.750.75 on Scientific and Instruments and 11 elsewhere.
  • (3) Principled training design. Fixing t=1t=1 and anchoring rr at the all-mask state aligns training with the one-step query. Within our method SPRINT, the instantaneous-style setting ρ=0\rho=0 is never the best, and larger ρ\rho also converges faster (refer to Figure 3 in Page 10, and Figure 7 in Page 31 in our updated version).

Weakness 2 & Question 4: Isolating the token-level objective from SID-level matching and ranking: "Could you provide a control that keeps the SID branch and inference procedure unchanged while replacing the token-level contrastive objective with ordinary per-token cross-entropy?"

Thank you for this constructive suggestion. To address your concern, I have design a variant:

  • V1: replacing the token-level contrastive objective with ordinary per-token cross-entropy while keeping others unchanged.

Besides, we also design two more variants to comprehensively evaluate the effectivess of our contrastive formulation:

  • V2: replacing the SID-level contrastive objective with ordinary cross-entropy while keeping others unchanged.
  • V3: replacing both token-level and SID-level contrastive objective with cross-entropy.

As shown in the Table 3 below, we have the following finding.

Key Finding: Removing either side contrastive objective harms model performance to a great extent. This directly verifies the effectiveness of our contrastive formulation.

Table 3. Performance comparison between different variants in terms of contrastive formulation.

MethodSportsBeautyToysScientificInstrumentsGames
R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10R@10N@10
SPRINT (ours)0.06210.03470.09870.05520.09980.05940.06560.03690.07130.03860.10010.0534
V10.04660.02470.08210.04630.09000.05470.04820.02800.05150.02800.07450.0408
V20.04420.02340.08310.04600.08860.05440.04320.02480.04740.02590.06960.0374
V30.02830.01550.06450.03810.07560.04770.03070.01870.03240.01880.05370.0310
AuthorsOpenReview ↗

Responses to Reviewer ZMtL (Part 3)

Weakness 3 & Question 5: Efficiency comparison protocol: "What comparison set does “second-fastest AR/NAR method” refer to, and how should the abstract’s speedup claim be interpreted in light of the faster SID-MLP baseline in Table 10?" “In addition, Appendix D.3 states that efficiency comparisons use standardized configurations, whereas accuracy comparisons use separately tuned configurations...” “Eq. (11) also requires scoring valid SIDs, so candidate scoring and top-(K) selection costs matter in addition to the single network pass. Although Appendix F acknowledges the dependence on catalog size, the timing boundaries and the inclusion of these costs need to be explicit.”

(1). Efficiency comparison statement.

The efficiency comparison in the abstract and Introduction is made against accuracy-oriented AR and NAR GR methods, i.e., TIGER, ActionPiece, RPG, LLaDA-Rec, DiffGRM, and OneGR. SID-MLP belongs to a different category, efficiency-oriented GR methods, together with AtSpeed, NEZHA, and EARN. We compare SPRINT with this category separately in Table 10 (Page 29). SPRINT is also the fastest there, 1.661.66 to 2.23×2.23\times faster than the fastest efficiency-oriented baseline SID-MLP (refer to Table 10 in Page 29 in our updated version), while being substantially more accurate (e.g., R@5 0.0444 vs. 0.0295 on Scientific).

To avoid ambiguity, we have revised the corresponding statements in the Abstract and Introduction of the updated version as follows:

"8.39−10.04×8.39-10.04\times speedup over the fastest accuracy-oriented GR baselines, and 1.66−2.23×1.66-2.23\times speedup over the fastest efficiency-oriented GR baselines"

(2). Why a standardized configuration, and why the comparison is fair and robust.

Latency depends directly on batch size, sequence length, and beam width. If each method kept its own settings, the measured gaps would mix the effect of the method with the effect of these settings. We therefore fix them for all methods in the efficiency comparison (batch sizes 1024/256/256, sequence length 20, beam width 10 where applicable). This choice gives SPRINT no extra advantage. On the contrary, it makes our comparison more conservative, as explained below.

  • The original best settings of baseline methods would only enlarge our efficiency advantage.

The best settings of beam-based baselines typically use a beam width of 50 or even larger one. Their decoding cost grows with the beam width, so their latency would increase substantially. SPRINT uses no beam search, so its speedup would be even larger. The standardized setting is thus a conservative choice for us.

  • For the fastest accuracy-oriented baseline, the standardized setting is its best setting.

RPG uses a beam width of 10 in its original paper, which is exactly our setting. The headline speedup of 8.398.39 to 10.04×10.04\times is therefore measured against RPG at its own optimal configuration.

(3). Timing boundaries.

To make the cost of each stage explicit, we further decompose SPRINT's inference time into the network pass and the scoring plus top-KK selection (as shown Table 4 below). We run each dataset five times and report the mean, with the minimum and maximum in parentheses.

Table 4: Breakdown of SPRINT's inference time (ms/sample; mean (min to max)).

Dataset#ItemsNetwork passScoring + top-KKTotal
Toys11,9242.972 (2.926 to 2.987)0.576 (0.571 to 0.583)3.548
Beauty12,1012.991 (2.908 to 3.153)0.581 (0.570 to 0.583)3.572
Sports18,3572.959 (2.938 to 2.989)0.590 (0.583 to 0.593)3.549
Instruments24,5872.959 (2.931 to 2.991)0.607 (0.607 to 0.614)3.567
Games25,6122.963 (2.937 to 2.982)0.605 (0.600 to 0.608)3.568
Scientific25,8482.972 (2.968 to 3.011)0.606 (0.603 to 0.613)3.578

Findings.

Finding 1: The network pass is independent of the number of items.

It stays between 2.959 and 2.991 on all six datasets, although the number of items more than doubles (11,924 to 25,848). The small differences show no trend with the number of items and fall within the run-to-run range of each dataset, so they reflect normal fluctuations from hardware and system factors. SPRINT runs one forward pass per user, regardless of the catalog size.

Finding 2: Scoring and top-KK take a small share of the time and grow slowly with the number of items.

They account for only 16.2% to 17.0% of the total. As the number of items more than doubles, their cost rises only from 0.576 to 0.606.

Together, these findings show that the network pass dominates the inference cost, while the catalog-dependent part stays small. Reducing the number of network passes is therefore the key to efficiency, which is exactly what single-step generation achieves. Multi-step AR and NAR methods pay this dominant cost LL times, further multiplied by the beam width.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.