acceptodds
Under review as a conference paper at ICLR 2027

DIFFUSION GENERATIVE RETRIEVAL OF ORDERED SLATES

Abstract

In generative retrieval, the model memorizes the index in its own parameters. In this setting, the recommender stores the catalog in its weights and emits each item directly as a tuple of semantic ID tokens, without an inference time candidate pool. Diffusion is a natural decoder when the entries of a list are chosen jointly. Earlier diffusion approaches operate along the identifier levels of a single item and assemble the slate from separate decodes. Diffusion Generative Retrieval of Ordered Slates (DiffROS) transposes this axis. It denoises a canvas that holds the full slate, with a slot for every position. At each identifier level, DiffROS predicts every slot in the context of all others. The standard offline objectives, however, are modular set functions. They score each item separately and add those scores across the slate. Ranking items by their marginals, or separate contributions, therefore already maximizes these objectives. Conditioned on retrieval, DiffROS orders the recovered items several times more sharply than beam search and the autoregressive slate decoder. A next item decoder read out by beam search recovers more of the target items, while DiffROS better orders the ones it recovers. At matched capacity and with a slate of one hundred, we found that DiffROS recovers close to three times as many targets as an autoregressive slate decoder while its decoding depth remains constant in slate size, with three forward passes rather than three hundred.

Primary area:
generative models
Keywords:
Generative Retrieval, Semantic ID, Masked Diffusion, Recommender Systems, Slate Recommenders, Generative Rankers, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_S6qYOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

DiffROS generates ordered slates of semantic IDs by resolving identifier levels sequentially and slate positions in parallel, with optional duplicate repair. It compares this factorization with item-level beam retrieval and autoregressive slate generation, and studies duplicate accounting and conditional ordering metrics. The reported gains concern decoding depth and ordering diagnostics; item-level beam retrieval remains stronger on the internal-corpus retrieval metrics.

Critical strengths
  1. Reduced sequential decoding depth. Parallelizing slate positions makes the decoding depth independent of slate size at a fixed denoising and repair budget. This is a useful computational property, distinct from constant total computation.

  2. An informative duplicate-accounting check. Table 4 rescores the same outputs under two conventions, illustrating how repeatedly crediting a correct item can distort the comparison between repetitive and diverse predictions.

Critical weaknesses
  1. The central ordering advantage is not established by the proposed diagnostic. Sections 4.1 and 4.4 compare gaps between two halves of each decoder's own recovered items under a deliberately bimodal target. This mixes ordering with the composition of the recovered set: even two perfectly ordered sets can have different half-split gaps if their relevance ranges differ. A decoder retrieving consistently relevant items can have a smaller gap than one retrieving a mixture of high- and low-relevance items. The shuffle control demonstrates an ordering signal within DiffROS's recovered set, but does not establish superior ordering across different recovered sets.

  2. The practical comparison relies on a severely repetitive AR baseline. In Table 3, the autoregressive decoder returns only a few distinct identifiers even for long slates. The matched-capacity and unrepaired-versus-unrepaired comparisons are useful, but do not establish that this behavior is inherent to autoregression. An AR baseline with duplicate prevention or repair, including its actual computational cost, would make the practical comparison more convincing. The public-benchmark results are mixed, so they do not independently establish a consistent advantage. Missing internal-corpus split details and decoder training specifications also make the severe repetition difficult to assess independently.

  3. The claimed exactness of the reverse process is not established. Section 2 calls the level-wise sampler an “exact reverse transition,” while Appendix C states that the reverse process factorizes. Equation (5) gives coordinate marginals, but independent forward masking does not imply conditional independence of dependent slots revealed together. For example, a distribution supported equally on (0,1) and (1,0) has uniform marginals whose independent product also generates (0,0) and (1,1).

I recommend rejection. The computational property and duplicate-accounting analysis have value, but the comparative evidence does not establish the proposed joint decoder's main benefit, and the reverse-process justification needs correction.

Minor Issues

  • I recommend a dedicated related-work section to organize and substantively discuss these important connections, rather than relying on the current introduction alone. In particular, distinguish DiffROS from DDBC [1] for masked discrete diffusion over RVQ codes in multi-item construction; HiGR [2] for efficient semantic-ID slate generation through hierarchical planning; GFN4Rec [3] for autoregressive list generation with reward-aligned diversity; and List-CVAE [4] for direct joint slate generation.

    [1] Tu, T., Li, A., Ma, Y., Xu, S., Liu, X., Ma, H., Pang, L., & Chua, T.-S. (2026). Discrete diffusion for bundle construction. The Fourteenth International Conference on Learning Representations.

    [2] Pang, Y., Liu, Z., Li, Y., Zhu, S., Luo, Z., Yu, C., Wu, S., Shen, S., Xu, C., Wang, B., Jiang, K., Yu, H., Zhuo, C., & Li, Z. (2026). HiGR: Efficient generative slate recommendation via hierarchical planning and multi-objective preference alignment (arXiv:2512.24787v2). arXiv.

    [3] Liu, S., Cai, Q., He, Z., Sun, B., McAuley, J., Zheng, D., Jiang, P., & Gai, K. (2023). Generative flow network for listwise recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 1524–1534). Association for Computing Machinery.

    [4] Jiang, R., Gowal, S., Qian, Y., Mann, T. A., & Rezende, D. J. (2019). Beyond greedy ranking: Slate optimization via List-CVAE. International Conference on Learning Representations.

  • Section 4.3's “cannot reveal an arrangement” is too broad: nDCG responds to order, though it does not measure inter-item utility.

  • important Section 4.2 and the conclusion claim beam's advantage on both corpora, but Table 2 has no Amazon beam row. =

  • Figure 3's timing panels need hardware, batch size, precision, and caching details

Questions for authors
  1. Separating ordering from item selection (W1). How does the half-split gap distinguish better ordering from a wider relevance range or a different recovered item population? Could you evaluate ordering on a matched item set and include a pointwise scoring-and-ordering control? Please also explain why the within-decoder subtraction in Appendix E is claimed to cancel the selection issue when those differences are subsequently compared across decoders.

  2. A practical AR comparison and the value of joint decoding (W2). Could you compare against an AR slate decoder with duplicate prevention or repair, reporting both retrieval/ordering quality and the full added decoding cost under comparable conditions? Beyond the existing factorization comparisons, what evidence shows when bidirectional joint conditioning improves the ordered-slate objective, rather than benefiting mainly from the severe repetition of the tested AR baseline?

  3. Conditions for an exact reverse transition (W3). Under what assumptions do the coordinate-wise reverse probabilities in Equation (5) yield an exact joint transition for slots revealed together? Could you distinguish exact sampling from the factorized model defined by Equation (4) from recovery of the data distribution's true reverse process? Please also clarify how Appendix D's uniform reopening is consistent with a transition that otherwise leaves unmasked positions fixed, or identify it separately as a repair operation.

5 comments

Thank you for the thorough review and for crediting the computational property and the duplicate-accounting check.

Separating ordering from selection (W1, Q1).\textbf{Separating ordering from selection (W1, Q1).} Evaluated on units both decoders emit, the control you requested, DiffROS orders the two halves of the target better than beam on shared first-level codes. Both decoders are scored on the same pairs of shared units (target identifiers or first-level codes that both emit); on first-level codes, cross-half pair accuracy is about 0.10 higher with and without repair, with 95% intervals excluding zero, and at exact identifiers the comparison is inconclusive (Sec. 4.4, p. 9; App. G, Table G2, p. 20). Beam over one item serves as the pointwise scoring-and-ordering baseline, since it receives no slot input; the first-level advantage is consistent with slot-position conditioning, and we did not run a control that isolates cross-slot conditioning or evaluate duplicate-aware autoregressive ordering at thirty slots. Table 1's gaps are now described as within-decoder, and Appendix E no longer compares their magnitudes across decoders (Table 1 caption, p. 6; App. E, p. 18).

Practical comparison and full cost (W2, Q2).\textbf{Practical comparison and full cost (W2, Q2).} Against an autoregressive decoder with duplicate prevention, the baseline you asked for, DiffROS with repair keeps the lead across the full slate, and we report its added computation. It reaches 1.28×1.28\times [1.26, 1.31] the recall@100 and 1.23×1.23\times [1.22, 1.25] the hit@100 of masked autoregressive decoding with 12 instead of 300 sequential passes; the price is about 12×12\times the position evaluations, since each pass processes the whole canvas, and the masked decoder leads at the top ten (Sec. 4.2, p. 7, ll. 370-377; p. 8, l. 397). Because the repaired margin persists against an autoregressive decoder that cannot repeat an identifier, the remaining lead does not rest on the original baseline's repetition. Decode time for the new configurations was not measured (App. G, p. 19, ll. 1010-1012). Section 4.1 now gives the split and training protocol, identical across decoders: one pass over the same training window, the final checkpoint, a held-out evaluation day, and no valid-ID constraint (Sec. 4.1, p. 6, ll. 290-294).

Exactness of the reverse process (W3, Q3).\textbf{Exactness of the reverse process (W3, Q3).} You are right that independent forward masking does not make the joint reverse transition exact, and the revision says so. Eq. 4 is now a stated modeling assumption: at one step per level and without repair, the sampler draws exactly from that factorized model, which equals the data's reverse conditional only when the slots revealed together are conditionally independent, exactly what your (0,1)/(1,0) example breaks (Sec. 2, p. 4, ll. 205-209 and ll. 214-215; App. C, p. 15, ll. 795-797). Reopening resolved positions is not a transition of Eq. 5, so Appendix D treats every reopening rule, uniform reopening included, as a decode-time repair operator (App. D, p. 17, ll. 897-900). No measured result depends on exactness.

Related work.\textbf{Related work.} The Introduction now distinguishes DiffROS from each work you named, within the page limit. DDBC completes partially observed, unordered bundles by masked diffusion over residual codes; HiGR plans a slate autoregressively and then beam-decodes each item, so its depth grows with slate size; GFN4Rec selects items one at a time from a candidate set; List-CVAE decodes a short slate in one pass and resolves each slot to a catalog item. DiffROS combines a request-conditioned canvas of positioned slots, residual identifiers decoded without a candidate pool, and depth that does not grow with slate size (Introduction, p. 2, ll. 65-74). We placed this discussion in the Introduction's literature review rather than a separate section to stay within ten pages.

Minor issues.\textbf{Minor issues.} We addressed the minor issues you raised. Section 4.3 now says that nDCG reflects arrangement through its rank weights (p. 8, l. 431); the beam comparison now refers to the internal corpus; Section 4.2 now gives the decode benchmark's batch size, precision and caching, and states that it times an untrained model on a synthetic catalogue, so it measures computational structure rather than deployment latency (p. 8, ll. 404-407); Figure 2's caption lists its configuration differences (p. 7, ll. 344-348); and, as Section 4.2 already states, the public benchmark is mixed at three items and a single-item soundness check at one (Table 2).

Please Note: page and line pointers in our first-round replies refer to the previous revision, available under Revisions; the latest revision adds one paragraph to Section 2, which moves later content down by about one page.

Reviewer_sVJPOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DiffROS, a diffusion-based generative retrieval framework for generating ordered slates of semantic IDs. It enables parallel generation across slate positions while preserving sequential decoding across identifier levels. The paper also analyzes the limitations of standard retrieval metrics in evaluating slate generation. Experimental results show that DiffROS can better capture slate ordering.

Critical strengths
  1. Generating multiple items as an ordered slate is interesting for GR. The proposed approach preserves the hierarchical dependencies of semantic IDs while allowing parallel generation across items, offering a reasonable design for reducing sequential decoding costs.

  2. The paper explains why recall and nDCG may not fully capture dependencies between items, and shows how duplicate predictions can affect evaluation results. These observations are useful for understanding and evaluating generative slate retrieval.

  3. The experiments provide useful insights into the behavior of different decoding strategies. The authors also discuss cases where DiffROS performs worse than beam search and acknowledge several limitations of the evaluation.

Critical weaknesses
  1. In Table 3, the autoregressive slate baseline generates only about 2.3 distinct identifiers for a slate of 100 items, which seems unusually low. The paper explains that this baseline uses greedy decoding without preventing duplicate generation. However, it is unclear how much of the reported improvement comes from this decoding strategy. A comparison with duplicate-aware autoregressive decoding would make the results more convincing.

  2. The paper explains why parallel generation across items is reasonable. But it is less clear why diffusion is particularly suitable for this design. As shown in Equation (4), the default decoding process resembles level-wise parallel prediction, which could also be implemented without diffusion. A comparison with a conventional level-wise parallel decoder would help clarify the specific contribution of diffusion.

Questions for authors
  1. How would the results change if the autoregressive slate baseline used duplicate-aware decoding?

  2. What benefits does diffusion provide beyond the proposed level-wise parallel factorization? Would a non-diffusion decoder with the same factorization achieve similar results?

  3. Is the reference for Chen and Karger (2006) correct? I found the following citation:

@inproceedings{chen2006less,
  title={Less is more: probabilistic models for retrieving fewer relevant documents},
  author={Chen, Harr and Karger, David R},
  booktitle={Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval},
  pages={429--436},
  year={2006}
}
3 comments

Thank you for finding ordered slate generation an interesting problem and for crediting the modularity analysis, the insight into decoding strategies, and the candor about where beam wins.

Duplicate-aware autoregressive decoding (W1, Q1).\textbf{Duplicate-aware autoregressive decoding (W1, Q1).} Duplicate-aware decoding strengthens the autoregressive baseline considerably, yet DiffROS with repair still leads it across the full slate with 25 times fewer sequential passes. To your question of how much of the improvement comes from the decoding strategy: without duplicate prevention in either decoder, DiffROS has 3.32×3.32\times the autoregressive recall@100 in the same evaluation; with duplicate handling in both, the margin is 1.28×1.28\times [1.26, 1.31] in recall@100 and 1.23×1.23\times [1.22, 1.25] in hit@100, at 12 instead of 300 sequential passes (Sec. 4.2, p. 7, ll. 370-377 and p. 8, l. 397; App. G, p. 19). The masked autoregressive decoder leads at the top ten and overtakes DiffROS without repair, which the revised paper reports alongside.

What diffusion adds (W2, Q2).\textbf{What diffusion adds (W2, Q2).} What our design adds in practice is decode-time repair from the same checkpoint. The network is trained on partially revealed slates, so it can reopen a duplicated slot and predict it again in the context of the rest of the slate while excluding the duplicated identifier; three repair rounds raise recall@100 2.32×2.32\times over no repair (Table G1). We did not train a level-wise parallel decoder without diffusion masking, so we did not test whether such a decoder could support the same repair; we present repair as the practical benefit of our design rather than as a proven advantage over it (Limitations, p. 10, ll. 524-526).

Chen and Karger (Q3).\textbf{Chen and Karger (Q3).} Thank you for catching the Chen & Karger reference; the citation and our paraphrase are corrected (Sec. 3, p. 5, ll. 239-241).

Thank you for the detailed response and additional experiments. The duplicate-aware autoregressive baseline largely addresses my first concern. The new results make the comparison more convincing.

My main remaining concern is the motivation for using diffusion. Since diffusion is central to the paper's claimed contribution, I consider this an important issue. The authors have shown that iterative repair is effective. However, the paper still lacks a convincing justification for why a diffusion framework is needed, rather than a conventional level-wise parallel decoder. The rationale behind this design choice remains unclear. I would be willing to raise my score if this concern is adequately addressed.

Thank you for the follow-up, and for noting that the duplicate-aware baseline largely addresses your first concern. We have added a short version of this rationale to Section 2 (p. 5, ll. 227-235); page and line pointers refer to the revision uploaded just now.

Why masked diffusion rather than a level-wise parallel decoder.\textbf{Why masked diffusion rather than a level-wise parallel decoder.} What we take from the diffusion framework is training on partially revealed canvases, which gives one network learned predictions for a slot given other slots at its own level, not only given the committed lower levels. A level-wise parallel decoder trained with teacher forcing sees every slot revealed below level ℓ\ell and every slot masked from ℓ\ell upward, so it has no trained conditional of this kind; reassigning a duplicated slot must then rely on its original per-slot predictions or a hand-designed rule. Our training views mask nested subsets of the most difficult canvas positions (Sec. 2, p. 3, ll. 145-153), and the predictor needs no noise-level input (p. 3, ll. 136-138), so one checkpoint serves both one pass per level and repair. Training a level-wise decoder on such partially revealed levels would give it the same capability, and that is the masked-denoising training we use.

The evidence.\textbf{The evidence.} At one step per level and without repair, DiffROS has the level-wise parallel factorization, with slots conditionally independent within a level given the lower levels (Sec. 2, p. 5, ll. 216-220), and in that configuration the duplicate-aware autoregressive decoder beats it on recall@100 by 1.81×1.81\times. With repair, which reopens each duplicated slot and re-predicts it given the realized slate with the duplicate excluded, DiffROS leads that decoder at cutoff one hundred, by 1.28×1.28\times in recall@100 and 1.23×1.23\times in hit@100, while the autoregressive decoder still leads in the top ten (App. G, Table G1, p. 20). This shows that repair works on our checkpoint; whether a fixed reassignment rule on a level-wise decoder could match it is not tested, and we now list that comparison in the Limitations (p. 10, ll. 538-539).

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

DiffROS directly generates ordered semantic-ID slates without a request-specific candidate pool. It preserves residual identifier levels sequentially while predicting slate slots in parallel, using coherent noising and optional duplicate repair. The paper analyzes modular ranking objectives and separates retrieval coverage from conditional ordering. On natural internal targets with 100 slots, unrepaired DiffROS achieves approximately 2.96 times the autoregressive baseline's recall using 3 rather than 300 sequential forward calls. However, the reported internal comparisons favor single-item beam search for recall. Public Amazon experiments cover only one- and three-item outputs.

Critical strengths
  1. Controlled comparisons and transparent reporting. The internal experiments share the tokenizer, encoder, training window, and nearly identical parameter counts. The paper acknowledges beam's recall advantage and the distinction between teacher agreement and user preference.
  2. A concrete evaluation finding. In Table 4, crediting each target only once changes AR's nDCG from 0.21917 to 0.05000, while beam remains at 0.14037. This demonstrates a consequential evaluation issue for these generated outputs.
Critical weaknesses
  1. The methodological increment over prior work is insufficiently isolated, and the ordering evidence does not establish the claimed mechanism. MaskGR already demonstrates parallel generation of multiple items represented by semantic IDs, SoundStorm combines coarse-to-fine residual-codebook decoding with parallel prediction across positions, and the OCN training scheme is explicitly inherited from DiffGRM. Moreover, UniRank's w/o LPS ablation (§3.2, Table 2) directly generates SID tokens over the entire vocabulary, so comparison only with its candidate-constrained main model does not fully establish the distinction. Possible differences in candidate-pool conditioning and residual-level scheduling should be clarified. The manuscript acknowledges several of these connections, and its commit-scope ablations provide useful evidence. However, the methodological contribution beyond adapting and combining these established components remains insufficiently established.

    The ordering diagnostic also does not isolate the contribution of cross-slot conditioning. The half-split comparisons use different recovered sets: a model recovering widely separated target ranks can obtain a larger gap than another perfectly ordering a narrower subset. Although Appendix E acknowledges confounding in absolute relevance levels, recovered-set composition also affects the magnitudes and ratios of the half-split gaps. Shuffling each model's outputs establishes nonrandom order within its own recovered set, but does not establish comparable ordering capability across models. For the unrepaired DiffROS rows (r=0r=0) in Table 1, the first-level codes are conditionally independent given the request under Equation (4) at T=1T=1. Position-specific marginals may therefore explain that result. A matched comparison with the closest multi-item masked-SID approach, a position-aware control without cross-slot interaction, and an ordering evaluation that controls recovered-set composition would help establish the incremental methodological contribution and its proposed mechanism.

  2. The practical advantage is not established against a reasonable noncollapsing AR control. At 100 slots, Table 3 reports only 2.32 distinct SIDs for AR. Appendix D explains the AR collapse through argmax decoding, whereas Section 2 explicitly describes categorical sampling for DiffROS. Please clarify the actual evaluation policies; without this clarification, matched capacity alone does not establish that the comparison isolates factorization. Valid-ID and no-repeat constraints, matched sampling, and usable-item quality–cost comparisons are missing. Repair also lowers validity to 0.7260 at this length. Figure 3 does include timing microbenchmarks; the unresolved issue is delivery cost for valid, distinct items, not an absence of timing measurements.

Questions for authors
  1. Additional evidence needed for current claims: Does the ordering advantage survive a control retaining slot positions but removing cross-slot interaction, and evaluation on common recovered items? Please report first-level and complete-SID results with uncertainty. This tests the present mechanism claim; an online deployment study is future work, not required here.
  2. Additional evidence needed for current claims: With matched sampling and valid-ID/no-repeat constraints, what are AR and DiffROS's recall, valid distinct-item counts, and measured decoding costs at 30 and 100 slots? This tests the existing factorization and efficiency claims, rather than requiring a new industrial system.
  3. Clarification: Please clarify the assumptions underlying the reverse-process and likelihood interpretations in Equations (4)–(8). Is the slot-factorized reverse distribution a modeling assumption, or is it intended to represent the exact reverse conditional? How is reopening committed positions consistent with Equation (5), which keeps resolved positions fixed? Appendix C acknowledges that the implemented training objective differs from the likelihood bound; how should this distinction be reflected in Appendix D’s interpretation of the position-wise loss as a sequence-NLL term? If any statements require qualification or revision, please explain which main conclusions depend on them and which remain supported.
6 comments

Thank you for the detailed review and for crediting the controlled comparisons, the open reporting of beam's lead, and the duplicate-accounting result.

Evaluation policy.\textbf{Evaluation policy.} Every number in the submitted paper comes from deterministic argmax decoding in both slate decoders, so the original comparison was policy-matched; Section 2 now states this explicitly (Sec. 2, p. 4, ll. 214-215). The sentence that suggested sampling described an available readout, not the one we evaluated.

Matched-sampling and duplicate-aware controls.\textbf{Matched-sampling and duplicate-aware controls.} In both controls we ran, matched sampling and duplicate handling in both decoders, DiffROS keeps the full-slate recall lead over the autoregressive decoder. When both decoders sample the first identifier level, DiffROS reaches 3.70×3.70\times the autoregressive recall@100; with duplicate handling in both, repaired DiffROS reaches 1.28×1.28\times [1.26, 1.31] the recall@100 and 1.23×1.23\times [1.22, 1.25] the hit@100 at 12 instead of 300 sequential passes, about 12 times the position evaluations (Sec. 4.2, p. 7, ll. 374-376; App. G, Table G1 and paired ratios, p. 19). Masked autoregressive decoding leads at the top ten and overtakes DiffROS without repair, and we did not run sampled DiffROS with repair. We sample only the first level because sampling deeper levels of a multi-level identifier compounds invalid identifiers.

Validity and usable items.\textbf{Validity and usable items.} Validity is reported for every new decoder in Table G1; because it is pooled over distinct identifiers, we report it descriptively rather than as a ranking. Repair does lower pooled validity, as you note, but recall and hit rate credit only distinct, real target items, so the margins above count usable recoveries. We evaluate every decoder unconstrained on the internal corpus, the open-catalog setting we study, where validity and duplication are measured rather than enforced (Sec. 2, p. 2, ll. 98-101; Sec. 4.1, p. 6, l. 294); duplicate-aware autoregressive decoding at thirty slots, catalog-resolved counts of valid distinct items and decode time for the new configurations remain future work, so passes and position evaluations describe computational structure rather than delivery time (App. G, p. 19, ll. 1010-1012).

Ordering with recovered-set composition controlled.\textbf{Ordering with recovered-set composition controlled.} On units both decoders emit, the evaluation you asked for, DiffROS orders the two halves of the target better than beam on shared first-level codes, consistent with slot-position conditioning. Both decoders are scored on the same pairs of shared units; on first-level codes, cross-half pair accuracy is about 0.10 higher with and without repair (95% intervals exclude zero), while at exact identifiers and against the autoregressive decoder the comparison is inconclusive (Sec. 4.4, p. 9; App. G, Table G2, p. 20). We did not train a position-aware control without cross-slot attention, so we no longer attribute the effect to cross-slot interaction.

Theory.\textbf{Theory.} We corrected the theoretical wording you flagged, and none of the measured results depends on it. Eq. 4 is now stated as a modeling assumption: at one step per level and without repair, the sampler draws exactly from this factorized model, which equals the data's reverse conditional only when slots revealed together are conditionally independent (Sec. 2, p. 4, ll. 205-209 and ll. 214-215; App. C, p. 15, ll. 795-797). Reopening a resolved position is not a transition of Eq. 5, so Appendix D now presents every reopening rule, uniform reopening included, as a decode-time repair operator, and calls the position-wise training loss a surrogate for the identity-assignment likelihood rather than a bound on it (App. D, p. 16, ll. 850-854; p. 17, ll. 897-900). Depth, recall, duplicate accounting and ordering are direct measurements and are unaffected by this correction.

Contribution relative to prior work.\textbf{Contribution relative to prior work.} DiffROS's contribution is now stated precisely against MaskGR, SoundStorm and UniRank. MaskGR is a next-item method: it unmasks the semantic-ID tokens of one item in parallel. DiffROS instead decodes ordered slates of up to one hundred positioned items, committing one residual level across all slots per pass and adding duplicate-keyed repair (Introduction, p. 1, l. 53 - p. 2, l. 55), a regime where duplication becomes a first-order problem. Section 2 credits SoundStorm for the coarse-to-fine schedule (p. 3, ll. 116-117), and, as in the submission, Appendix C credits DiffGRM for the coherent noising we train with (p. 16, ll. 819-820). UniRank's main model scores a supplied candidate pool, and its ablation without latent pool selection, as we read its input definition (history, candidate pool and slate in one sequence, UniRank Sec. 2.2), still receives that pool; DiffROS conditions only on the request. We did not run a matched comparison with MaskGR; the closest evidence is the commit-scope ablation in Section 2, where level-scoped commits outperform item-scoped and global commits. The priority claim is removed.

Public benchmark.\textbf{Public benchmark.} The public benchmark is a single-item soundness check plus small slates by construction, and the slate-size questions are answered on the internal corpus. Its slates are each user's last purchases, which were never displayed together, and training needs at least 3N+1 interactions per user, which limits the slate size (Sec. 4.1).

Thank you for the additional experiments and clarifications. The duplicate-aware AR baseline, shared-unit ordering evaluation, and theoretical corrections address important parts of my review. I have two remaining points.

1. Comparison with MaskGR. Your response describes MaskGR as a single-item method, but its Section 4.4 and Figure 4 include parallel two-item prediction, which your original Introduction also acknowledges. Could you reconcile these descriptions and make the comparison explicitly against that multi-item extension? Please clarify the incremental contribution relative to this setting, distinguishing the architectural differences from what the commit-scope ablation within DiffROS establishes.

2. Practical quality–cost comparison. The reported Recall@100 gain against no-repeat AR is meaningful. However, the reduction from 300 to 12 sequential passes comes with about 12 times the position evaluations. I appreciate your acknowledgment that decoding time and catalog-resolved valid-distinct counts for the new configurations remain unmeasured.

Could you report measured decoding time and the number of valid, distinct catalog items per request for repaired DiffROS and no-repeat AR at 100 slots, using matched hardware and batch size? These remain relevant to my original request. If unavailable, I would regard the current evidence as supporting reduced sequential depth and improved full-slate recall, while practical efficiency remains unresolved.

Comments on the paper

1 comment

We would like to sincerely thank the reviewers for their time and careful reading. The revision adds the evaluations you asked for, a duplicate-aware autoregressive baseline (all three of you) and an ordering evaluation on units both decoders emit (ymsh and S6qY), and each confirms a specific advantage of DiffROS: with three repair rounds it keeps the full-slate lead over the strongest of the four autoregressive decodings we evaluated, and on shared first-level codes it orders the target halves better than beam search. Textual additions are marked in green in the revised PDF, and the page and line pointers below refer to it.

Duplicate-aware autoregressive baseline.\textbf{Duplicate-aware autoregressive baseline.} With duplicate handling in both decoders, DiffROS with three repair rounds still leads the strongest of the four autoregressive decodings we evaluated across the full slate, using 25 times fewer sequential passes. In an additional evaluation of the same checkpoints, against autoregressive decoding with a last-level no-repeat mask, DiffROS with three repair rounds reaches 1.28×1.28\times [1.26, 1.31] the recall@100 and 1.23×1.23\times [1.22, 1.25] the hit@100 (paired request-bootstrap 95% intervals), at 12 instead of 300 sequential passes and about 12 times the position evaluations (Sec. 4.2, p. 7, ll. 370-376; App. G, p. 19). Without duplicate prevention in either decoder, the margin in that evaluation is 3.32×3.32\times.

The trade-off.\textbf{The trade-off.} The no-repeat mask makes the autoregressive decoder much stronger, enough to lead at the top ten and to overtake DiffROS without repair, and we report this trade-off in full; across the full slate, however, repaired DiffROS leads the masked autoregressive decoder on both recall@100 and hit@100, and without repair, DiffROS with a sampled first level reaches the highest hit@100 in this evaluation in only three passes (1.27×1.27\times the masked decoder's, at lower recall@100; App. G, Table G1, p. 19).

Ordering on shared units.\textbf{Ordering on shared units.} Evaluated on units both decoders emit, as requested, DiffROS orders the two halves of the target better than beam search on shared first-level codes. Cross-half pair accuracy is about 0.10 higher with and without repair, and the 95% intervals exclude zero; at exact identifiers, and against the autoregressive decoder, the comparison is inconclusive (Sec. 4.4, p. 9; App. G, Table G2, p. 20).

Corrections.\textbf{Corrections.} We also addressed each point the reviewers flagged; every textual addition is marked in green in the revised PDF. Section 2 now states that all reported slate decodes are argmax unless sampling is stated and that Eq. 4 is a modeling assumption; Appendix C limits exactness to the factorized model, and Appendix D treats repair as a decode-time operator and the training loss as a surrogate; Table 1's gaps are scoped to each decoder's own items; the Introduction positions DiffROS against MaskGR, UniRank and the works S6qY named; and the Chen & Karger citation, both figure captions (Figure 3's timing axis now correctly reads per batch of eight) and the training protocol (Sec. 4.1) are corrected or added.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.