acceptodds
Under review as a conference paper at ICLR 2027

From Next-Item to Multi-User Generation: Reformulating Cold-Start Recommendation via Parallel Generative Discrete Diffusion

Abstract

Most existing recommender systems are built from a user-centric perspective, aiming to recommend the most suitable content to each user. However, this paradigm often fails in cold-start scenarios, where the paucity of interaction feedback compromises both reliable representation learning and sufficient initial exposure within massive candidate pools. This paper advocates a paradigm shift from ”Next-Item” to ”Multi-User Generation” after revisiting the underlying logic of cold-start recommendation, and reformulates the task as **Item-centric Multi-User Generation**, aiming to proactively identify seed users to break the passive exposure deadlock. However, this transition is challenging, as the multi-modal target user distributions and the non-sequential nature of interactions pose a severe inductive bias mismatch for traditional user-centric methods. To address this challenge, we propose **P**arallel Generative **D**iscrete **D**iffusion (), which employs a permutation-invariant Masked Discrete Diffusion mechanism to construct a direct mapping from item content to the latent user space, while eschewing autoregressive constraints. We also develop Interval-Aware Spatio-Temporal Encoding, incorporating temporal biasing with positional encoding to capture the temporal evolution of group interests and ensure multi-user extrapolation. With this design, empowers effective *Multi-Group Look-ahead Prediction* to explore diverse interest groups and *Scalable Parallel Retrieval* to ensure high-efficiency decoding. Extensive experiments on two public real-world datasets and large-scale online A/B tests demonstrate that outperforms existing methods, especially in cold-start scenarios.

Primary area:
generative models
Keywords:
Generative Cold-Start Recommendation, Next-User Modeling, Mask Diffusion Model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.80 over 5 reviews, above 85% of the others.
Reviewer_gXVYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PD2, a paradigm shift from user-centric next-item prediction to item-centric multi-user generation, with the main use case being item cold-start. PD2 simultaneously generates multiple users from a given item via a non-autoregressive masked denoising process. Experiments on KuaiRand datasets demonstrated superior performance, especially for the cold-start setting.

Critical strengths
  1. The paper proposes a paradigm shift from U2I to I2U recommendation, which alleviates the bias against high-quality new items in the recommender system. Specifically, generating a pool of users given an item. This formulation addresses the practically important cold-start problem.
  2. The empirical experiment provides overall strong evidence for the effectiveness of the formulation and method. The paper evaluates cold, warm, and overall performance with other baseline methods. The results disentangled the benefits from problem formulation and detailed architecture design. The reported online engagement gains are practically meaningful.
Critical weaknesses
  1. Gap between theoretical analysis and practical experiments. The masked diffusion process is an over-statement of the deployed setting, which uses one-pass factorized retrieval rather than the iterative reverse process used in the theoretical proof. The architecture is not really permutation-invariant as claimed. This makes it unclear whether PD2 is a joint diffusion/set generator or a useful but more incremental multi-query masked retriever.
  2. Important information such as related work and online experiment setting/results are not properly discussed in the main text, only in the Appendix, which limits the coherency and completeness of the paper.
Questions for authors
  1. How were the user-to-item baselines adapted to the I2U setting for the baseline models? Can the authors provide more details to ensure it’s a fair comparison?
  2. In what precise sense is PD2 permutation-invariant and a joint diffusion generator, given its chronological positional encoding, temporal bias, factorized output channels, and one-pass inference? Have the authors compared the performance of PD2 with a non-diffusion multi-query retriever?
  3. Why KuaiRand datasets? Have the authors consider other datasets?
2 comments
AuthorsOpenReview ↗

Response to Reviewer gXVY - Part 1/2 (W1 & Q2 & W2)

We are grateful for your positive feedback and the thoughtful evaluation of our work. Please see below for our response.

(W1 & Q2) The reviewer argue taht there is a gap between the theoretical analysis and the practical experiments. Reviewer requests a comparison with multi-query retrieval.

R: We would like to clarify that PD² adopts Masked Discrete Diffusion for conditional denoising training, while employing Single-pass Retrieval for efficient multi-user inference. Its main contribution lies in task-specific mechanisms designed to address the modeling mismatch between existing I2U approaches and the concurrent, multi-modal characteristics of item-side interactions.

First, our adoption of Single-pass Retrieval is a deliberate design choice. Appendix A analyzes the idealized Masked Discrete Diffusion objective and continuous-time reverse process. Under appropriate assumptions, the practical training objective supports the learning of optimal conditional predictions. However, Single-pass Retrieval directly generates multiple user-retrieval queries in one forward pass rather than executing an iterative reverse chain. In practical scenarios, Single-pass Retrieval trades a small amount of retrieval performance for reduced computational overhead. To illustrate this trade-off, we provide an additional comparison between two inference strategies on KuaiRand.

Inference StrategyCold R@10Warm R@10All R@10
Single-pass Retrieval3.463.613.60
Iterative Denoising3.393.633.62

We have included the above comparative experiment in the revised manuscript (Lines 1227–1233).

Second, regarding permutation-invariant, We acknowledge that our claim of permutation invariance was overly broad, as incorporating RoPE introduces a trade-off between permutation invariance and positional modeling. However, narrowing the scope of this claim does not affect the conclusion of Theorem 1(ii). We have discussed this issue in detail. Please refer to our response to Reviewer Gs6w, W2 & Q1.

Third, comparison with Multi-query Retriever. We further construct a carefully adapted BERT-style Masked Retrieval baseline and progressively introduce parallel prediction and our task-specific components. The results on KuaiRand-Pure are shown below:

VariantR@10Diversity
(a)Adapted Masked Retrieval Model3.280.479
(b)(a) + Parallel Generation (10 Inference Tokens)3.100.548
(c)(b) + MGLP3.380.572
(d)(c) + IAST3.600.575

Simply extending Masked Retrieval to multiple parallel queries improves Diversity from 0.479 to 0.548 but reduces R@10 from 3.28 to 3.10. Introducing MGLP improves R@10 to 3.38, while incorporating IAST further increases R@10 to 3.60 and Diversity to 0.575. These results demonstrate that the improvements of PD² are not solely attributable to multiple queries, but also depend on the complementary mechanisms designed for I2U multi-user retrieval.

Finally, we provide additional experiments comparing PD² with two groups of non-diffusion multi-query retrievers.

We use SASRec and TIGER as base models and adapt them to support multi-query retrieval through two strategies:

  1. Autoregressive: The prediction from the previous step is fed back as input for the next step, ultimately generating 10 predictions.
  2. MultiQuery: Following MIND [1], we introduce 10 learnable Query Embeddings, which aggregate historical representations through attention to produce 10 distinct prediction vectors for multi-query retrieval.

We report the comparison results on KuaiRand-Pure below:

MethodR@10N@10
SASRec (Base)1.961.07
SASRec (Autoregressive)1.320.71
SASRec (MultiQuery)1.871.05
TIGER (Base)2.341.55
TIGER (Autoregressive)1.900.98
TIGER (MultiQuery)2.211.47
PD² (Ours)3.601.85

These results highlight the importance of complementary mechanisms specifically tailored to I2U multi-user retrieval.

[1] Li, Chao, et al. "Multi-interest network with dynamic routing for recommendation at Tmall." ​CIKM 2019.


(W2) The reviewers pointed out that important information such as related work and online experiment setting/results are not properly discussed in the main text, only in the Appendix.

R: Thank you for raising this point. Due to space limitations, we originally placed some discussions of related work and online A/B test results in the Appendix. To improve the coherence and completeness of the paper, we have moved the related work and the key results from the online A/B tests to the main text (Section 2 and Lines 474-475).

AuthorsOpenReview ↗

Response to Reviewer gXVY - Part 2/2 (Q1 & Q3)

(Q1) The reviewers would like to know about the adaptation of U2I baselines to the I2U setting.

R: Thank you for your attention to baseline adaptation and the fairness of our comparisons. We would like to clarify that, to fairly compare the modeling capabilities of different methods for I2U cold-start recommendation, we adapt methods originally designed for U2I to the I2U setting while preserving their original architectures and core training mechanisms as much as possible.

First, we unify the task inputs and prediction targets. Traditional U2I methods typically take a user's historical item interaction sequence as input and predict the next item. In the I2U setting, we transform this into an item-centric historical user interaction sequence. Specifically, we construct a chronologically ordered user sequence based on each item's positive interaction records and predict the future interacting user. Accordingly, the prediction space is changed from the item vocabulary to the user vocabulary.

Second, we adapt different baselines according to their original modeling mechanisms. For sequential modeling methods, we preserve their original attention structures and sequence modeling mechanisms while adapting the inputs and prediction targets to item-centric user sequences. For generative methods such as DreamRec and DiffuRec, we change the generation target from future item representations to future user representations. For semantic-ID-based methods, we follow the standard practice of using Qwen3-4B to obtain multimodal embeddings and RQ-KMeans to construct semantic IDs, while retaining the next-token prediction task.

Finally, we adopt a unified experimental protocol to ensure fair comparisons. All methods use the same data splits, full user candidate vocabulary, and Recall@K/NDCG@K evaluation metrics. Furthermore, in Appendix C.3 (Table 5), we compare the same architectures under both U2I and I2U training formulations to distinguish the effects of changing the task formulation from those of the model architecture itself.

We have added detailed adaptation procedures for each baseline in Appendix B.3 (6) of the revised manuscript to improve experimental transparency and reproducibility.


(Q3) The reviewer would like to know why the KuaiRand datasets were selected.

R: Thank you for raising this question. KuaiRand dataset is one of the most widely used public datasets in the recommendation community. Its two public subsets provide complementary evaluation settings from different perspectives, and their scales exceed those of many existing public recommendation benchmarks.

Furthermore, to evaluate the generalization ability of beyond the short-video domain, we provide experiments on two Amazon dataset. (Appendix Table 4)

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

This paper focuses on cold-start recommendation, where new items suffer from limited interaction feedback and insufficient initial exposure under the conventional user-to-item paradigm. The authors reformulate the task as item-centric multi-user generation and propose Parallel Generative Discrete Diffusion (PD²). Specifically, the method adopts masked discrete diffusion to predict multiple potential users in parallel, and introduces interval-aware spatio-temporal encoding and multi-group look-ahead prediction to capture diverse and evolving user interests. Offline experiments and online A/B tests demonstrate the effectiveness of the proposed framework.

Critical strengths

S1. The paper presents a well-motivated and coherent connection between the cold-start problem, the I2U formulation, and discrete diffusion modeling. Proactively identifying seed users directly addresses the exposure difficulty of new items, while parallel masked generation naturally fits the concurrent and multi-modal nature of item-side interactions. In my view, this alignment between the practical problem, task formulation, and modeling choice is particularly compelling.

S2. The paper provides informative visual evidence supporting the effectiveness of PD². In particular, Fig. 3 first compares the same SASRec backbone under Next-Item and Next-User modeling, and then shows the additional benefit of PD². Together with the MMD results, these analyses provide representation-level evidence for both the I2U formulation and the proposed modeling approach, beyond the improvements in recommendation metrics.

Critical weaknesses

W1. Some important experimental settings are not sufficiently described. In particular, the data splitting procedure, the definition of test targets, and the candidate user set need clarification. For zero-interaction items, the paper should also specify the available content features and whether all compared methods use the same information. These details are important for reproducing and interpreting the cold-start results.

W2. The description of multi-channel retrieval is inconsistent between the main text and Appendix E.2. Eqs. (10)–(11) retrieve approximately K/N users from each channel and take their union, which may yield fewer than K users due to duplicates. However, Appendix E.2 retrieves Top-K users from each channel and then merges them into a final Top-K list. It is unclear which strategy is used in the reported experiments.

Questions for authors

Q1. Could the authors clarify how the training, validation, and test sets are constructed? Are the test targets the next interacting user or all users within a future window, and is retrieval evaluated over all users or sampled candidates? What content features are available for zero-interaction items?

Q2. Which multi-channel retrieval strategy is used in the accuracy, diversity, and efficiency experiments? How are duplicate users handled to obtain the final Top-K list?

Q3. The connection between the theoretical analysis in Appendix B and single-pass inference is unclear to me. Could the authors explain which theoretical guarantees apply to the implemented method?

4 comments
AuthorsOpenReview ↗

Response to Reviewer bYAk - Part 1/2 (W1 & Q1)

We are grateful for your positive feedback and the thoughtful evaluation of our work. Please see below for our response.


(W1 & Q1) The reviewer would like to know more specific details about the experimental setup.

R: Thank you for raising this point.

Data splits: We follow the widely adopted chronological leave-one-out protocol in sequential recommendation [1] and generative recommendation studies [2, 3]. Specifically, for each item, we sort its positive interactions chronologically, reserve the penultimate interaction for validation and the final interaction for testing, and use all preceding interactions for training.

Test targets & candidate user set: For offline evaluation, we use a chronological leave-one-out protocol, using the last user to interact with each item as the true label. All methods use the same full user vocabulary of each dataset as the candidate pool, rather than a sampled subset, ensuring a consistent evaluation protocol. For online serving, we restrict the candidate pool to recently active users, as described in Appendix B.4 (Line 1077).

Features available for zero-interaction items: Each item is represented using item-side metadata, such as multiple tags, video duration, and music type, which are encoded and fused through a shared feature encoder. All compared methods have access to the same item-side features, ensuring a fair comparison under identical information availability, including for zero-interaction items.

We have added supplementary experimental details in Appendix B.3

[1] Kang, Wang-Cheng, and Julian McAuley. "Self-attentive sequential recommendation." ​*2018 IEEE international conference on data mining (ICDM)*​. IEEE, 2018.

[2] Hou, Yupeng, et al. "Generating long semantic ids in parallel for recommendation." ​Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2​. 2025.

[3] Liu, Zhanyu, et al. "Onerec-think: In-text reasoning for generative recommendation." arXiv preprint arXiv:2510.11639 (2025).

AuthorsOpenReview ↗

Response to Reviewer bYAk - Part 2/2 (W2 & Q2 & Q3)

(W2 & Q2) The reviewer asks which multi-channel retrieval strategy is used and how duplicate users are handled to obtain the final Top-K list.

R: Eqs. (10)–(11) in the main text present the basic idea of uniformly allocating retrieval quotas across multiple parallel channels. However, since different channels may retrieve the same users, directly merging their results may lead to wasted retrieval quotas. To avoid this issue, we adopt cross-channel deduplication and global merging in practice, thereby constructing a final recommendation list containing KK distinct users. We have added an explanation of this procedure in the main text (Lines 342–345).

To illustrate this with a simple example, when retrieving the final Top-10 users, we first retrieve the Top-10 users from each parallel prediction channel. We then merge and deduplicate the candidate users across all channels. If the same user appears in multiple channels, we retain their highest ranking. Finally, we select the global Top-10 users from the deduplicated candidate set.


(Q3) The reviewers would like to know which theoretical guarantees apply to this implementation method.

R: Thanks for your valuable comments. The theoretical guarantees applicable to PD² mainly concern conditional prediction consistency. Specifically, (1) absorbing-state discrete diffusion ensures that masked-token reconstruction remains a categorical prediction problem over the finite user vocabulary. (2) under realizability and population-risk optimality, positive time reweighting preserves the optimal conditional predictions, while NCE can recover the corresponding conditional distributions under appropriate negative-sampling and density-ratio assumptions

Moreover, our adoption of Single-pass Retrieval is a deliberate design choice. In practical scenarios, Single-pass Retrieval trades a small amount of retrieval performance for reduced computational overhead. To illustrate this trade-off, we provide an additional comparison between two inference strategies on KuaiRand.

Inference StrategyCold R@10Warm R@10All R@10
Single-pass Retrieval3.463.613.60
Iterative Denoising3.393.633.62

We have included the above comparative experiment in the revised manuscript (Lines 1227–1233).

The author has addressed my main concerns, and I decide to keep my positive score.

Thank you for your response. We are glad that we have addressed your concerns.

Reviewer_NAE3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues for replacing the user-centric "next-item" (U2I) formulation of cold-start recommendation with an item-centric "multi-user generation" (I2U) one: given an item, generate the users likely to engage with it, so that new items can be matched to seed users instead of waiting to be retrieved. PD2 tokenizes an item's history as [item token, user tokens ...], trains a bidirectional Transformer with an absorbing-mask discrete-diffusion objective (BERT-style masking at rate t, NCE against popularity-sampled negatives), and at inference appends N mask tokens whose predictions serve as N parallel retrieval channels. An interval-aware temporal attention bias combined with RoPE, and a window-averaged ("multi-group look-ahead") NCE target, support this. Validation: two KuaiRand datasets against nine I2U-adapted baselines, ablations, two Amazon categories, latency measurements, and a 7-day online A/B test on a 400M-DAU platform comparing U2I vs. proactive I2U serving.

Critical strengths

S1. The reformulation is properly controlled. App. D.3 is the right experiment: the same backbones (SASRec, BERT4Rec, HSTU, PAM, NU) re-trained under both U2I and I2U on a common retrieval pipeline show that item-centric supervision helps cold items for every architecture, while PD2 still wins under a fixed formulation. Together with the cold/warm MMD analysis (Fig. 3), this separates "the formulation helps" from "our model helps" — something most paradigm-shift papers never check.

S2. The online evidence is the paper's strongest asset. A/A pre-check, user-level randomisation, 5% traffic, 95% CIs and p-values, and gains that grow as item exposure falls (+15.6% LTR for videos with <200 plays) directly support the claim the paper is actually making. Deployment evidence at this scale is rare in the cold-start literature.

S3. The systems design is deployable and the efficiency story is concrete. Dual-mode serving (standard U2I retrieval over PD2 item embeddings, plus asynchronous I2U push into an ANN index of active users) integrates a next-user model without disrupting an existing stack; parallel decoding gives 1.9×/4.6×/9.5× at 2/5/10 tokens (Table 8). Code is released.

Critical weaknesses

W1. The objective that is proved is not the objective that is trained, and the theory contradicts the method. Eq. (23) defines the NELBO with a 1/t weight; Eq. (6), the actual loss, has no such weight, and substitutes NCE for the categorical likelihood. Since E[|M_t|] = tL, the unweighted sum down-weights exactly the high-noise regime the ELBO is derived to balance. Worse, Theorem 1(ii) rests on exchangeability (Assumption 1), while the whole point of the interval-aware bias — worth +6.1% R@10 in Table 3 — is that order and timing do carry information. Either the data is exchangeable, in which case IAST is pointless, or it is not, in which case the optimality claim does not apply. The window-averaged target (Eq. 16) also moves the objective outside the theorem entirely. Inference compounds this: the N masks are filled in one forward pass with no iterative denoising, so what is deployed is a one-shot masked autoencoder with N query slots, not a diffusion sampler.

W2. Training and inference see structurally different inputs. During training, masked positions lie inside 1..L, so every masked token has observed tokens to its right; at inference the N masks are appended with no right context at all. The paper frames this only as "unseen positional indices", but RoPE cannot fix a missing context pattern. A suffix-masking training variant is the obvious control and is absent, so the central mechanism is only partly validated (the "w/o Bi-Attention" ablation does not test it, since it still trains with the masked objective).

W3. The cold-start result on zero-history items is unexplained, and it is the central claim. KuaiRand-1K has only 1,000 users and 4.37M items, so an item with no training interactions receives no gradient on its token e_v; yet Fig. 7 shows PD2 on 0-history items at roughly the level of the most popular bin. Since Sec. 2.1 mentions multi-field features but the model uses only a single item token (no content encoder is described), it is unclear whether the signal comes from static content, from an item embedding trained on data overlapping the evaluation window, or both. A shuffled-item-ID control and a content-only cold-start baseline are needed; otherwise the comparison is against interaction-based baselines on a task that may be solved by content.

W4. The paradigm claim is not isolated from the architecture. Every I2U baseline is discriminative while PD2 is generative, and the only diffusion-for-recommendation baseline (LLaDA-Rec) is cited but not run, so "masked diffusion beats autoregression" is never tested. The "mirrored I2U" adaptation is one sentence for models (TIGER, DreamRec/DiffuRec, HSTU) where it is far from trivial. There are no standard deviations despite 5 runs. Minor: the "first / pioneer" claims should be softened (next-user retrieval exists — NU is a baseline — as does masked diffusion for recommendation); "permutation-invariant" is inaccurate given RoPE and timestamp-dependent bias; the online I2U-vs-U2I gap (+67.9% LTR) is attributed to allocation policy while candidate pool, exposure accounting and impression counting also change, and "same exploration budget" is never defined.

Questions for authors
  1. What is actually optimised, and at which mask pattern? Is the 1/t weight of Eq. (23) applied in code, or is Eq. (6) as written? Do you train with suffix-only masks (appended after L, as at inference) or only interior masks — and how much Recall is recovered if you switch? Given that NCE (Eq. 7) and the window-averaged target (Eq. 16) are not covered by Theorem 1, which of the Appendix-B guarantees do you claim for the implemented model?
  2. Exchangeability vs. the temporal bias. Assumption 1 requires the user sequence to be exchangeable, yet IAST relies on non-exchangeable interval structure and contributes +6.1% R@10. Which claim about the data is intended, and does Theorem 1(ii) survive dropping Assumption 1?
  3. Zero-history items. Where does the signal come from in Fig. 7, given that the model's only item input is e_v? Are the item features fully static content, or do they include interaction statistics, and is any training data for these items drawn from the evaluation window? Please report a shuffled-item-ID control and a content-only baseline.
  4. Baselines and variance. Give the exact I2U adaptation recipe for TIGER, DreamRec, DiffuRec and HSTU, and confirm equal tuning budgets. Can you add (a) an autoregressive generative next-user baseline and (b) LLaDA-Rec-style masked diffusion under I2U, plus standard deviations/significance tests for Table 2? Also, define "the same exploration budget" for the online I2U-vs-U2I comparison and report any guardrail metrics.
3 comments
AuthorsOpenReview ↗

Response to Reviewer NAE3 - Part 1/3 (W1/W2 & Q1/Q2)

We are grateful for your positive feedback and the thoughtful evaluation of our work. Please see below for our response.

(W1/W2 & Q1/Q2) The reviewer argues that the objective analyzed in the theoretical proof differs from the actual training objective, and that the theory contradicts the proposed method.

R: We thank the reviewer for highlighting the distinction between the theoretical objective and the practical implementation. We address it from the following three folds:

(1) Training Objective and Its Theoretical Connection. The actual training objective is the MGLP loss in Eq. (16), which extends the masked reconstruction objective in Eq. (6) by introducing window-averaged supervision and employs the NCE loss in Eq. (7). We sample a diffusion time t∼U(0,1)t \sim U(0,1) and independently mask user tokens with probability tt. No explicit 1/t weighting is applied during training. We acknowledge that this practical objective is not identical to the continuous-time NELBO analyzed in Theorem 1. Nevertheless, the two objectives share the same optimal conditional predictions. We provide the detailed explanation of the 1/t weighting, NCE, and MGLP in our response to Reviewer Gs6w (W2 & Q2).

(2) Exchangeability and Temporal Modeling. First, IAST does not inherently contradict permutation invariance. Its interval-aware temporal bias depends on the actual time differences between interactions rather than their positional indices. Therefore, jointly permuting interaction tokens and their corresponding timestamps preserves the temporal relationships and the permutation-equivariant property of temporal attention. However, there is a trade-off between strict permutation invariance and RoPE, which introduces positional information to capture weak sequential dependencies and support length extrapolation.

Importantly, removing the exchangeability assumption does not affect the theoretical conclusion. (Theorem 1(ii)) We have revised the proof in Appendix A to establish the consistency of the ideal iterative reverse process through value-independent random unmasking and the chain rule, without requiring exchangeability. Thus, the conclusion of Theorem 1(ii) remains valid under the revised assumptions. We provide more detailed discussion in the response to Reviewer Gs6w (W2 & Q1).

(3) Training Mask Patterns and Inference. During training, we follow the standard BERT-style random masking strategy adopted in masked discrete diffusion, where user tokens are independently masked with probability t∼U(0,1)t \sim U(0,1). This allows the model to learn conditional predictions under diverse masking patterns and corruption levels, rather than being restricted to suffix-only masks. During inference, we append NN future mask tokens and predict their corresponding user representations in parallel, enabling efficient multi-user retrieval in a single forward pass. Our adoption of Single-pass Retrieval is a deliberate design choice. In practical scenarios, Single-pass Retrieval trades a small amount of retrieval performance for reduced computational overhead. To illustrate this trade-off, we provide an additional comparison between two inference strategies on KuaiRand.

Inference StrategyCold R@10Warm R@10All R@10
Single-pass Retrieval3.463.613.60
Iterative Denoising3.393.633.62

We have included the above comparative experiment in the revised manuscript (Lines 1227–1233).

Finally, we provide two additional control experiments on KuaiRand-Pure, comparing Suffix-only Masking and the removal of Bidirectional Attention (As shown in Table 3).

MethodR@10N@10
PD² (Suffix-only Masking)3.561.83
PD² (w/o Bi-Attention)3.421.76
PD² (Full)3.601.85

The results show that random masking achieves slightly better performance than suffix-only masking, supporting its effectiveness in learning from diverse masking patterns. Moreover, the performance drop without Bidirectional Attention further demonstrates its importance for modeling global dependencies among user interactions.

AuthorsOpenReview ↗

Response to Reviewer NAE3 - Part 2/3 (W3 & Q3)

(W3 & Q3) The reviewer points out that the strong performance of PD² on zero-history items remains insufficiently explained and seeks clarification on the sources of item representations and additional control experiments.

R: Thank you for this insightful comment. We would first like to clarify that the item token does not contain only ID information, and all methods use the same item features. Each item is represented using item-side metadata, such as multiple tags, video duration, and music type, which are encoded and fused through a shared feature encoder. These features do not include interaction statistics such as play counts or click-through rates.

Importantly, the feature encoder is shared across items and learns associations between item content and user interests from the training interactions of other items. Therefore, even when an item has no historical interactions, PD² can generate its representation from the available content features and identify potential users. This cross-item generalization provides an explanation for the strong zero-history performance observed in Fig. 7.

We have added a clarification of the available item features in Appendix B.3 (3).

Regarding Data Splitting and Information Leakage: We follow the chronological evaluation protocol and define zero-history items based on their training interaction histories. For these items, predictions should rely exclusively on content features available at inference time, without using interactions or interaction-derived statistics from the evaluation window.

Finally, we provide two additional control experiments on KuaiRand-Pure, reporting R@10:

MethodZero-history ItemsCold-start ItemsWarm Items
PD² (Shuffled-item-ID)3.323.363.39
PD² (Content-only)3.373.433.50
PD² (Full)3.373.463.61

These results demonstrate that PD² can effectively generalize to zero-history items by learning transferable item–user associations. Meanwhile, the superior performance of full PD² on warm items further highlights its ability to leverage historical interactions for more accurate user retrieval.

AuthorsOpenReview ↗

Response to Reviewer NAE3 - Part 3/3 (W4 & Q4)

(W4 & Q4) The reviewer argues that the contribution of the I2U paradigm has not been sufficiently disentangled from the model architecture and seeks further clarification on the experimental details and fairness.

R: Thank you for your attention to the contributions of the paradigm and model architecture, as well as experimental details.

(1) Disentangling the Contributions of the Paradigm and Architecture: First, at the representation level, Fig. 3 analyzes the Cold–Warm representation gap under different training paradigms through visualization and MMD. Switching the same architecture to I2U training substantially reduces this gap, while PD² further improves alignment, as supported by the MMD analysis.

Second, at the quantitative evaluation level, in Appendix C.3 (Tab. 5), we train the same models under both U2I and I2U formulations and evaluate them using a unified standard U2I retrieval pipeline. The Cold R@10 of SASRec increases from 1.72 to 2.53, HSTU from 2.15 to 3.07, and PD² from 2.82 to 4.03.

Together, the representation analysis and quantitative experiments demonstrate that the I2U training paradigm itself improves representation learning for cold-start items, while the architectural mechanisms of PD² specifically designed for multi-user retrieval provide additional gains. Therefore, the performance improvements of PD² do not arise solely from changing the training paradigm, but from the combined effects of the paradigm and task-specific architectural design.

(2) I2U adaptation recipe: We adapt all baselines according to their original modeling mechanisms. For sequential modeling methods, we preserve their original attention structures and sequence modeling mechanisms while adapting the inputs and prediction targets to user sequences. For generative methods such as DreamRec and DiffuRec, we change the generation target from future item representations to future user representations. For semantic-ID-based methods, we follow the standard practice of using Qwen3-4B to obtain multimodal embeddings and RQ-KMeans to construct semantic IDs, while retaining the next-token prediction task.

All methods use the same data splits, full user candidate vocabulary, and evaluation metrics. We also adopt the same model dimensions and use Optuna for hyperparameter tuning, as described in Lines 1014–1016.

We have added detailed adaptation procedures for each baseline in Appendix B.3 (6) of the revised manuscript to improve experimental transparency.

(3) Additional Baselines and Variance: We have added a comparison with LLaDA-Rec to Table 2 in the revised manuscript. An autoregressive generative next-user baseline is already included (e.g., TIGER). Although NU is not a generative model, it also performs autoregressive next-user prediction.

Furthermore, we provide additional experiments comparing PD² with two groups of non-diffusion multi-query retrievers. Please refer to the final part of our response to Reviewer gXVY (W1 & Q2) for details.

We additionally report the mean ± standard deviation across five independent runs for PD² and the main baselines on both datasets. Please refer to our response to Reviewer Gs6w (W3 & Q3) (4) for details.

(4) Minor issue:

  1. The "first / pioneer" claims should be softened.

We acknowledge that prior work has explored Next-User Retrieval and Masked Diffusion. We would like to clarify that ​the novelty of our work lies in reformulating traditional single-user prediction as Generative Multi-User Modeling to address the concurrent and multi-modal characteristics of user interactions in I2U scenarios, and in designing complementary training and parallel retrieval mechanisms​. This is the key distinction between PD² and existing methods.

  1. About permutation-invariant.

First, Temporal Bias depends on the actual time differences between interactions rather than their positional indices in the input sequence, and therefore does not break permutation invariant when interactions and their corresponding timestamps are jointly permuted. RoPE introduces a trade-off with permutation invariance. Nevertheless, narrowing the scope of this claim does not affect the conclusion of Theorem 1(ii). We have updated the relevant proofs in Appendix A. Please refer to our detailed discussion in the response to Reviewer Gs6w (W2 & Q1).

  1. Define the same exploration budget.

Both experimental groups are drawn from the same user population, with users randomly assigned to the two groups to ensure a fair comparison. We compare the effectiveness of two retrieval channels, PD²-U2I and PD²-I2U. Both channels target the same eligible cold-start video pool (Video View < 50k), while employing different retrieval strategies. Consistent with conventional channel-level evaluation practices in industrial retrieval systems, we evaluate the effectiveness of each retrieval strategy based on its attributed exposures.

Reviewer_6zCkOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper formulates cold-start recommendation as an item proactively finding multiple groups of potential users. It proposes PD2, which uses an item prefix and historical users for masked training, adds a temporal bias, RoPE, and look-ahead-window supervision, and generates multiple user-retrieval vectors in parallel at inference. The authors provide offline benchmarks, cross-domain and training-direction controls, and online results for two serving configurations.

Critical strengths
  • Proactive retrieval of multiple audiences for new items addresses a meaningful practical problem. Table 2 and Appendices D.2–D.3 show that the gains are related to the training direction but are not explained entirely by changing that direction.
  • The empirical effort is substantial. The online experiments include user-level randomization, confidence intervals, and equal exploration budgets, and the appendices provide cross-domain, latency, and serving-mode comparisons. I regard these as important supporting evidence rather than judging the paper by its conceptual labels alone.
Critical weaknesses
  • The central claimed properties do not match the implemented algorithm. RoPE and directional windows do not guarantee permutation invariance. The appendix's analysis of a weighted categorical objective and continuous-time reverse process also does not directly establish that negative sampling, window-based supervision, and one-pass factorized retrieval recover a general joint user distribution. This affects the scope of the main contribution.

  • Details that determine the results remain unspecified. The offline temporal split, timestamps for future mask positions, target windows, and cold-item definition are unclear, while the main text and appendix describe different candidate-allocation and merging procedures. As written, accuracy, diversity, and NDCG cannot be verified under a single unambiguous protocol.

  • No required "AI use statement" section before Reference.

Questions for authors
  • Core theoretical clarification: What precisely is meant by permutation invariance here? How do the implemented Equations (6)–(7) and (16) relate to Appendix Equation (23), and why would a joint-consistency result for infinitesimal reverse steps apply to one-pass parallel retrieval? If these connections cannot be established, which central claims should be withdrawn or narrowed?
  • Protocol clarification that could change my assessment: Please specify the offline split, target windows, cold-item definition, and timestamp assignment for future mask positions, and identify whether the main results use the main-text quota allocation or the appendix's per-channel Top-K merging. Are the reported online rates computed over exploration exposures or all exposures? These answers would directly affect my recommendation.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 6zCk - Part 1/2 (W1 & Q1)

We sincerely appreciate the reviewer's helpful feedback and careful reading. Please see below for our response.

(W1 & Q1) The reviewer seeks clarification on the precise meaning of permutation invariance and the relationship between the actual training objective and the theoretical objective analyzed in the appendix.

R: Permutation Invariance mainly refers to the fact that, in I2U scenarios, user interactions with the same item within a short period of time typically do not exhibit causal dependencies that must follow their order in the interaction logs (as illustrated in Fig. 1). We further analyze the relationship between RoPE and permutation invariance, Please refer to Reviewer Gs6w's W2 & Q1.

Relationship between Eqs. (6)–(7), Eq. (16), and Appendix Eq. (23): Appendix Eq. (23) defines an original-position user reconstruction objective weighted by 1/t1/t; Eqs. (6)–(7) replace it with a negative-sampling objective without explicit time weighting; Eq. (16) further extends the original-position supervision target to user labels sampled from a look-ahead window.

The connection between our practical design and theoretical analysis is mainly reflected in the following two aspects:

  • Consistency of Optimal Conditional Predictions:

    Specifically, our training objective is derived from the bound in Theorem 1(i). The NELBO in Eq. (23) and the training objective in Eq. (6) differ by the removal of the 1/t factor and the replacement of cross-entropy with NCE (see (2)). Since 1/t is a positive constant independent of θ\theta for each fixed t, time reweighting only changes the weights of the non-negative KL terms associated with different contexts. Therefore, under the realizability assumption and at the population-risk optimum, a(t)=1/ta(t)=1/t and a(t)=1a(t)=1 yield the same optimal conditional predictions, and Theorem 1(ii) applies to Eq. (6) unchanged. This is consistent with the analysis of time-independent conditional distributions in absorbing discrete diffusion [1], and simplified time weighting also has precedents in prior work [2].

    For MGLP, Theorem 1 covers point-wise supervision, i.e., w=1w=1 in Eq. (16), which coincides with Eq. (6). For the look-ahead case w>1w>1 raised by the reviewer, the label of position kk is drawn uniformly from the masked users in Wk\mathcal W_k, so the objective targets the window-averaged conditional 1∣Wk∣∑j: euj∈WkP⋆(euj(0)∣Stv)\frac{1}{|\mathcal W_k|}\sum_{j:\ e_{u_j}\in\mathcal W_k}P^\star(e_{u_j}^{(0)}\mid\mathcal S_t^v) instead of the exact-position conditional of Eq. (30).

    This is a task-driven relaxation for I2U retrieval, which asks which users will interact in the near future rather than which log slot each of them occupies. We empirically validate the effectiveness of this relaxation through an ablation study on KuaiRand-Pure. Specifically, introducing MGLP into the adapted multi-query masked retriever improves R@10 from 3.10 to 3.38 and Diversity from 0.548 to 0.572. Further incorporating IAST increases R@10 to 3.60 and Diversity to 0.575. These results demonstrate the practical benefits of window-averaged supervision for multi-user retrieval. Our adoption of Single-pass Retrieval is a deliberate design choice. In practical scenarios, Single-pass Retrieval trades a small amount of retrieval performance for reduced computational overhead. To illustrate this trade-off, we provide an additional comparison between two inference strategies on KuaiRand.

  • Empirical Comparison:

    Our adoption of Single-pass Retrieval is a deliberate design choice. In practical scenarios, Single-pass Retrieval trades a small amount of retrieval performance for reduced computational overhead. To illustrate this trade-off, we provide an additional comparison between two inference strategies on KuaiRand.

Inference StrategyCold R@10Warm R@10All R@10
Single-pass Retrieval3.463.613.60
Iterative Denoising3.393.633.62

We have included the above comparative experiment in the revised manuscript (Lines 1227–1233).

[1] Ou et al. Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data. ICLR, 2025.

[2] Ho et al. Denoising Diffusion Probabilistic Models. NeurIPS, 2020.

[3] Ma and Collins. Noise Contrastive Estimation and Negative Sampling for Conditional Models: Consistency and Statistical Efficiency. EMNLP, 2018.


AuthorsOpenReview ↗

Response to Reviewer 6zCk - Part 2/2 (W2 & Q2 & W3)

(W2 & Q2) The reviewer seeks clarification on the experimental details.

Thanks for your valuable comments, we provide experimental details below.

Offline data split: We follow the widely adopted chronological leave-one-out protocol in sequential recommendation [1] and generative recommendation studies [2, 3]. Specifically, for each item, we sort its positive interactions chronologically, reserve the penultimate interaction for validation and the final interaction for testing, and use all preceding interactions for training.

Target Windows: During inference, the NN future mask positions serve as parallel user retrieval channels (Eq. 8). In our work, diversity improves as NN increases, therefore, we select N=10N=10 (Fig. 4). For offline evaluation, we adopt the chronological leave-one-out protocol, using the last user who interacted with each item as the ground-truth label.

Cold-item Definition: We define cold-start items as those whose historical interaction lengths fall within the bottom 10% (Appendix B.2.3). Furthermore, to evaluate performance under different degrees of cold-start severity, we provide a more fine-grained analysis in Appendix C.4 (Fig. 7), where items are divided into six groups according to their training interaction frequencies and evaluated separately.

Timestamp Assignment for Future Mask Positions: During inference, we assign the same timestamp to all future mask positions, using the timestamp of the most recently observed interaction plus a small offset. This places the future mask positions within the same local temporal window as the latest observed interaction, while avoiding the need to assign separate future timestamps to each prediction channel.

Quota Allocation: Eqs. (10)–(11) in the main text present the basic idea of uniformly allocating the retrieval quota across multiple parallel channels. However, since different channels may retrieve the same users, directly merging their results may lead to wasted retrieval quotas. To avoid this issue, we adopt cross-channel deduplication and global merging in actual retrieval, thereby constructing a final recommendation list containing KK distinct users. We have added an explanation of this procedure in the main text (Lines 342–345).

Online Evaluation Protocol: Following the online evaluation practice in industrial retrieval systems [4, 5], we report engagement rates over actual exposures attributed to each evaluated retrieval channel, rather than all exposures across the recommendation platform. Specifically, PD² and the online baseline model are evaluated under the same exploration budget, targeting videos with fewer than 50,000 views. Thus, the reported LTR, WTR, CMTR, and SVTR characterize channel-level engagement effectiveness.

[1] Kang, Wang-Cheng, and Julian McAuley. "Self-attentive sequential recommendation." ​*2018 IEEE international conference on data mining (ICDM)*​. IEEE, 2018.

[2] Hou, Yupeng, et al. "Generating long semantic ids in parallel for recommendation." ​Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2​. 2025.

[3] Liu, Zhanyu, et al. "Onerec-think: In-text reasoning for generative recommendation." arXiv preprint arXiv:2510.11639 (2025).

[4] Zhou, Xiaoyou, et al. "CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval." Proceedings of the 20th ACM Conference on Recommender Systems. 2026.

[5] Cao, Hannan, et al. "OneTrans-V2: Unifying Retrieval, Pre-rank, and Fine-rank with One Transformer in Industrial Recommender." arXiv preprint arXiv:2609.28589 (2026).

We have provided more detailed experimental settings in the revised manuscript (Appendix B.3).


(W3) The reviewer points out that the AI Use Statement section should be moved to immediately before the References.

R: Thank you for raising this point. We have moved the AI Use Statement section to immediately before the References.

Dear Reviewer,

Thank you very much for your time and for overseeing our submission.

We have carefully prepared a detailed, point-by-point response to your comments and concerns in the rebuttal.

As the discussion period is drawing to a close, we wanted to ask whether these answers resolve your questions. If you find the explanations adequate, we respectfully ask you to consider re-evaluating your rating. If any point would benefit from further analysis, or additional ablation, we would be glad to provide it while the discussion is open.

Your comments have helped us make our work more rigorous, and we sincerely appreciate your valuable feedback. Thanks again, we wish all the best in your work and life.

Best regards,

The authors

Reviewer_Gs6wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies item cold-start recommendation through item-to-user retrieval, where an item and its interaction history are used to identify potential seed users. The proposed PD2 framework combines a bidirectional masked-token model, interval-aware temporal attention biases, rotary positional embeddings, and a look-ahead supervision objective. At inference, multiple masked positions produce parallel retrieval queries whose candidate users are aggregated. Experiments report improvements on two KuaiRand datasets, two Amazon categories, and an industrial online A/B test.

Critical strengths
  1. The paper addresses a meaningful cold-start problem by proactively identifying seed users for new items. PD2 achieves substantial improvements over strong baselines, with additional results on Amazon supporting its applicability beyond short videos.
  2. The industrial A/B test provides valuable evidence of practical effectiveness. The comparison uses the same exploration budget and reports user-level randomization, confidence intervals, and statistically significant engagement improvements.
Critical weaknesses
  1. The framework mainly combines established components: bidirectional masked prediction, negative sampling, RoPE, and temporal attention biases. Since I2U recommendation already exists, the incremental contribution beyond a carefully adapted masked-retrieval model is insufficiently established.
  2. The claimed permutation invariance is difficult to reconcile with chronological inputs and position-dependent RoPE. Moreover, appendix b analyzes a weighted likelihood objective and continuous-time reverse process, whereas the actual method uses a look-ahead negative-sampling objective and single-pass retrieval. The theory therefore does not directly establish the claimed properties of PD2.
  3. The paper does not clearly describe the temporal splits, candidate user pools, or features available for zero-interaction items. These details are essential for assessing the validity and reproducibility of the cold-start results. Offline uncertainty estimates are also missing despite reporting five-run averages.
Questions for authors
  1. How is the claimed permutation invariance maintained when the model uses chronological inputs and position-dependent RoPE?
  2. Which theoretical guarantees apply to the actual look-ahead negative-sampling objective and single-pass retrieval procedure?
  3. Could you clarify the temporal splits, candidate user pools, and features available for zero-interaction items, and report uncertainty across the five runs?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Gs6w - Part 1/4 (W1)

Thanks a lot for your thoughtful comments. We present our responses below.


(W1) The reviewer argues that PD²'s incremental contribution over a carefully adapted masked-retrieval model is insufficiently demonstrated.

R: We would like to respectfully clarify that the core contribution of PD² is to address the modeling mismatch between existing I2U methods and the distinctive characteristics of item-side interaction distributions by proposing a task-driven generative modeling framework.

  1. Paradigm Level: Although I2U recommendation has been explored in prior work, existing methods mainly adopt single-target, sequential prediction paradigms by simply reversing the inputs and outputs of U2I models, overlooking the concurrent and multi-modal characteristics of item-side interactions. We are the first to reformulate I2U recommendation as Generative Multi-User Modeling, shifting from single-user prediction to generative modeling of the multi-user distribution within a future interaction window.
  2. Method Level: Rather than simply combining existing components, we develop a set of complementary mechanisms tailored to the unique characteristics of I2U. Specifically, Bidirectional Masked Diffusion alleviates unreasonable sequential dependencies, MGLP captures multi-modal user distributions through window-level multi-target supervision, while IAST and Parallel Retrieval capture dynamically evolving group interests and enable efficient, diverse user retrieval.
  3. Experimental Level: (1) Our existing ablation studies validate the individual contributions of each component. (2) We further introduce a carefully adapted BERT-style Masked I2U + Parallel Prediction + RoPE + Temporal Bias baseline to demonstrate the incremental benefits of our task-specific designs beyond general-purpose Masked Retrieval. As shown below, directly extending this model to multiple parallel queries increases Diversity from 0.479 to 0.548, but decreases R@10 from 3.28 to 3.10. Introducing MGLP further improves R@10 to 3.38, while incorporating IAST raises R@10 to 3.60 and Diversity to 0.575. These results demonstrate that simply masked retrieval model is insufficient to achieve both strong relevance and broad user coverage, highlighting the importance of MGLP and IAST specifically designed for the multi-user retrieval objective. (3) Figure 4 demonstrates that our method significantly improves user retrieval diversity while largely preserving accuracy, validating its effectiveness in supporting the Multi-User Modeling objective.
VariantR@10Diversity
(a)Adapted Masked Retrieval Model3.280.479
(b)(a) + Parallel Generation (10 Inference Tokens)3.100.548
(c)(b) + MGLP3.380.572
(d)(c) + IAST3.600.575

AuthorsOpenReview ↗

Response to Reviewer Gs6w - Part 2/4 (W2 & Q1)

(W2 & Q1)The reviewer asks how permutation invariance is maintained when the model uses chronological inputs and RoPE.

R: Thank you for raising this point.

First, chronological inputs are not inherently incompatible with permutation invariance. Without position-dependent encodings, a bidirectional transformer is inherently permutation-invariant.

Second, we further analyze the relationship between RoPE and permutation invariance below.

The I2U setting exhibits two distinctive characteristics:

  1. Short-term Concurrency: Concurrent interactions introduce uncertainty in the ordering of users within local windows, where multiple users may serve as valid prediction targets. By introducing local-window permutation invariance through MGLP, we reduce the model's dependence on the incidental ordering of concurrent interactions.
  2. Long-term Audience Evolution: As interactions accumulate, an item's target user groups may continuously expand and evolve over time (weak sequentiality). Particularly in cold-start scenarios, the model needs to extrapolate to user groups that have not been sufficiently observed. We introduce RoPE to incorporate relative positional structure, enhancing the modeling and generalization capabilities across different stages of interactions.

Although these two components may intuitively involve a certain trade-off, as local permutation invariance and global position awareness reflect different inductive biases, they are not directly contradictory. Instead, they address modeling requirements at different temporal scales: the former relaxes unnecessary ordering constraints in short-term concurrent interactions, while the latter preserves the temporal structure needed to capture long-term audience evolution.

Empirically, both components provide individual performance improvements in our ablation study, and their combination yields further gains.

VariantR@10N@10
PD² w/o MGLP & RoPE (Base)3.031.55
Base + MGLP3.521.80
Base + RoPE3.131.59
Base + Both3.601.85

From a theoretical perspective, we acknowledge that our claim of permutation invariance was overly broad, as incorporating RoPE introduces a trade-off between permutation invariance and positional modeling, as discussed above. However, narrowing the scope of this claim does not affect the conclusion of Theorem 1(ii). Due to space limit, we provide the proof sketch below. The full revised proof (Appendix A.1, Lines 868–912, highlighted in blue) relies on the chain rule and does not invoke exchangeability.

Under value-independent masking, Eq. (30) establishes single-token conditional matching for every mask pattern. In the continuous-time limit, masked positions are revealed sequentially in a value-independent, uniformly random order σ\sigma. For any fixed σ\sigma, the chain rule combines these conditionals into the same joint conditional distribution, and averaging over σ\sigma yields Eq. (33).

Let Rt\mathcal{R}_t and Mt\mathcal{M}_t denote the unmasked and masked positions of Stv\mathcal{S}_t^v, respectively, and let m=∣Mt∣m=|\mathcal{M}_t|. At the optimum of the masked objective, the model matches the true single-token conditional at every masked position (Eq.(29) in the paper):

=P^\star\bigl(e_{u_k}^{(0)}\mid\mathcal{S}_t^v\bigr), \qquad \forall\\ k\in\mathcal{M}_t. \qquad\text{(29)}$$ Since forward masking is independent of token values, the true posterior depends only on the revealed positions and their values, so Eq.(29) holds for every mask pattern. In the continuous-time limit, masked positions are revealed one at a time in a uniformly random order $\sigma$, independently of token values. Let $\mathcal{S}^{(i)}$ denote the state before the $i$-th reveal. For any fixed $\sigma$, we have $$\begin{aligned} &p_{\theta^\star}(\mathcal{S}_0^v \mid \mathcal{S}_t^v,\sigma)\\\\ &=\prod_{i=1}^{m} p_{\theta^\star}\bigl(e_{u_{\sigma(i)}}^{(0)}\mid\mathcal{S}^{(i)}\bigr)\\\\ &=\prod_{i=1}^{m} P^\star\bigl(e_{u_{\sigma(i)}}^{(0)}\mid\mathcal{S}^{(i)}\bigr)\\\\ &=P^\star(\mathcal{S}_0^v\mid\mathcal{S}_t^v), \qquad\text{(34-36)} \end{aligned}$$ where the second equality follows from Eq.(30) applied to each $\mathcal{S}^{(i)}$, and the last from the chain rule. Since (R2) holds for every $\sigma$, averaging over the revealing orders yields $$\begin{aligned} p_{\theta^\star}(\mathcal{S}_0^v\mid\mathcal{S}_t^v) &=\mathbb{E}_{\sigma}\bigl[p_{\theta^\star}(\mathcal{S}_0^v\mid\mathcal{S}_t^v,\sigma)\bigr]\\\\ &=P^\star(\mathcal{S}_0^v\mid\mathcal{S}_t^v). \qquad\text{(37)} \end{aligned}$$ **Therefore, Theorem 1(ii) remains valid without assuming exchangeability across time intervals.** ---
AuthorsOpenReview ↗

Response to Reviewer Gs6w - Part 3/4 (W2 & Q2)

(W2 & Q2) The reviewer seeks clarification on whether the theoretical analysis in the appendix applies to the actual training objective and single-pass parallel retrieval, and what theoretical guarantees can be established.


R: Thank you for pointing out the need to further clarify the connection between the theoretical analysis and the actual training objective. The connection between our practical design and theoretical analysis is mainly reflected in the following two aspects:

  • Consistency of Optimal Conditional Predictions:

    Specifically, our training objective is derived from the bound in Theorem 1(i). The NELBO in Eq. (23) and the training objective in Eq. (6) differ by the removal of the 1/t factor and the replacement of cross-entropy with NCE (see (2)). Since 1/t is a positive constant independent of θ\theta for each fixed t, time reweighting only changes the weights of the non-negative KL terms associated with different contexts. Therefore, under the realizability assumption and at the population-risk optimum, a(t)=1/ta(t)=1/t and a(t)=1a(t)=1 yield the same optimal conditional predictions, and Theorem 1(ii) applies to Eq. (6) unchanged. This is consistent with the analysis of time-independent conditional distributions in absorbing discrete diffusion [1], and simplified time weighting also has precedents in prior work [2].

    For MGLP, Theorem 1 covers point-wise supervision, i.e., w=1w=1 in Eq. (16), which coincides with Eq. (6). For the look-ahead case w>1w>1 raised by the reviewer, the label of position kk is drawn uniformly from the masked users in Wk\mathcal W_k, so the objective targets the window-averaged conditional 1∣Wk∣∑j: euj∈WkP⋆(euj(0)∣Stv)\frac{1}{|\mathcal W_k|}\sum_{j:\ e_{u_j}\in\mathcal W_k}P^\star(e_{u_j}^{(0)}\mid\mathcal S_t^v) instead of the exact-position conditional of Eq. (30).

    This is a task-driven relaxation for I2U retrieval, which asks which users will interact in the near future rather than which log slot each of them occupies. We empirically validate the effectiveness of this relaxation through an ablation study on KuaiRand-Pure. Specifically, introducing MGLP into the adapted multi-query masked retriever improves R@10 from 3.10 to 3.38 and Diversity from 0.548 to 0.572. Further incorporating IAST increases R@10 to 3.60 and Diversity to 0.575. These results demonstrate the practical benefits of window-averaged supervision for multi-user retrieval.

  • Empirical Comparison:

    Our adoption of Single-pass Retrieval is a deliberate design choice. In practical scenarios, Single-pass Retrieval trades a small amount of retrieval performance for reduced computational overhead. To illustrate this trade-off, we provide an additional comparison between two inference strategies on KuaiRand.

Inference StrategyCold R@10Warm R@10All R@10
Single-pass Retrieval3.463.613.60
Iterative Denoising3.393.633.62

We have included the above comparative experiment in the revised manuscript (Lines 1227–1233).

[1] Ou et al. Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data. ICLR, 2025.

[2] Ho et al. Denoising Diffusion Probabilistic Models. NeurIPS, 2020.

[3] Ma and Collins. Noise Contrastive Estimation and Negative Sampling for Conditional Models: Consistency and Statistical Efficiency. EMNLP, 2018.


Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Second Revision

We have uploaded a second revised version of our manuscript, incorporating the additional experimental results. Our claim remains unchanged: reformulating cold-start recommendation as item-centric multi-user generation, together with the task-specific design of PD², enables effective cold-start user retrieval and efficient parallel inference.

Since OpenReview's PDF diff compares the latest manuscript against the original submission, this comment summarizes only the changes made since the first version. All changes introduced in this second revision are highlighted in blue.

1. Revised Theoretical Proof (Appendix A.1). Since there is a trade-off between strict permutation invariance and RoPE, we have removed the exchangeability assumption from Theorem 1(ii). This does not affect the theoretical conclusion. Instead, we provide an alternative proof establishing the consistency of the ideal iterative reverse process through value-independent random unmasking and the chain rule, without requiring exchangeability of user interactions.

2. Improved Manuscript Coherence and Completeness. To improve the coherence and completeness of the manuscript, we have moved the Related Work section and the key online A/B testing results into the main text (Section 2 and Lines 474–475, respectively).

3. Additional Baseline and Inference Comparison (Table 2; Appendix B.2.2 and C.5). We have included LLaDA-Rec as an additional baseline. We also provide an empirical comparison between Single-pass Retrieval and Iterative Denoising, demonstrating comparable retrieval performance while avoiding iterative decoding.

4. Clarified Experimental Protocol and Baseline Adaptations (Appendix B.3). We have added details on chronological data splitting, full-vocabulary candidate evaluation, available item features, timestamp assignment for future mask positions, and the adaptation of U2I baselines to the I2U setting to improve reproducibility and experimental transparency.

5. Clarified Retrieval and Online Evaluation Procedures (Sections 3.2.3 and 4.2; Appendix B.3). We have clarified the cross-channel deduplication strategy and the attribution of online engagement metrics to the corresponding retrieval channels. We also highlight the online A/B findings in the main text, including the 10.48% LTR improvement of PD²-U2I over the production baseline and the further 67.91% improvement achieved by PD²-I2U over PD²-U2I.

6. Relocation of the AI Use Statement (Section 6). We have moved the AI Use Statement section to immediately before the References.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.