Listwise Preference Diffusion Optimization for User Behavior Trajectories Prediction

Thank you for your time and effort in reviewing our paper! We sincerely appreciate your thoughtful feedback and hope that our responses effectively address your concerns and support a score revision.

W1&Q1: The motivation of the new UBTP task is not well justified. In other words, why is the multi-step is better than the next-item prediction task?

Thank you for the question regarding the motivation behind introducing UBTP. While autoregressive next-item prediction can technically be extended to multi-step forecasting by rolling out one prediction at a time, we argue that this greedy, locally optimized generation strategy fails to capture global coherence and long-term dependencies inherent in user trajectories. We try to analyze it from four distinct perspectives:

• Cascading Error and Exposure Bias: Autoregressive rollout conditions each future prediction on previously generated items. Errors early in the sequence can propagate, degrading the quality of later predictions [1]. This is particularly problematic in recommendations, where early inaccurate suggestions can distort the entire predicted interaction path.

• Lack of Global Preference Modeling: Next-item models optimize local pointwise or pairwise objectives. In contrast, user trajectories are structured objects, where preferences emerge not only at the item level but also from inter-item dependencies, topic continuity, or progression (e.g., watching a trilogy or completing a course). These are not captured by locally greedy methods.

• Empirical Evidence: As shown in Tab.1, even strong autoregressive models like SASRec-ar underperform on long-sequence prediction metrics (e.g., SeqMatch), validating that non-autoregressive, trajectory-level modeling yields better alignment with real-world behavior.

• Non-Autoregressive Benefits: Our formulation enables parallel generation of full-length trajectories with listwise supervision, offering better sample diversity and efficiency compared to autoregressive alternatives.

Therefore, UBTP is not just a repackaging of next-item prediction, but a principled extension that acknowledges and addresses the limitations of autoregressive rollout when the goal is to model user behavior as a coherent, causally evolving sequence.

---

W2&Q2(i): Why not simply use the Top-1 Luce's Choice model?

Thank you for your thoughtful questions. The introduction of the Plackett–Luce (PL) model is a key design choice motivated by the need to capture global preference structures over full-length trajectories. While simpler alternatives like Top-1 Luce’s Choice model local preference signals, they fail to supervise the relative ordering among multiple relevant items.

Specifically, the PL model allows us to optimize the likelihood of the entire ordered item sequence simultaneously rather than focusing on the top-ranked item at each prediction step. This is essential because in our UBTP task, user preference across future steps exhibits significant ordering constraints and interdependencies. In contrast, the Top-1 model ignores the dependencies among subsequent positions beyond the top-ranked item [2-4], reducing its ability to produce coherent long-term user trajectories. Furthermore, the suggested Top-1 Luce's Choice model essentially corresponds to our baseline method, SASRec-AR [5], in the experiments, where predictions are optimized individually and independently at each step of the sequence.

Our experiments consistently show significant performance gains from the PL loss compared to the position-wise baseline methods (including SASRec-AR). LPDO achieves substantial improvements in all evaluation metrics, demonstrating that explicitly modeling the full-sequence ranking structure is necessary and beneficial for accurate multi-step user behavior prediction.

---

W2&Q2(ii): It seems that the LPDO method is dependent on K, i.e., the number of candidate items at each step. For different metrics, e.g., SeqNDCG@K, the K should also vary accordingly.

We would like to clarify that LPDO is not dependent on the number of candidate items at each step. During the optimization process, LPDO interacts with all candidate items for optimization without negative sampling. Specifically, the ListPref loss itself is formulated to consider the full candidate set, and its computation is independent of the evaluation metric.

Regarding evaluation metrics, following traditional sequential recommendation settings [5-7], the value of K in metrics such as SeqNDCG@K is determined by the evaluation protocol and does not affect the training procedure or the ListPref loss (Eq.11). In our experiments, we follow standard practice by reporting results for multiple values of K (e.g., K=5, 10, 20), and these are clearly specified in the experimental section.

---

W3&Q3: From my understanding, the next-item prediction is essentially a special case of UBTP with k=1. Since this task is more widely studied, the authors should also report the performance of next-item prediction in the experiments.

Thanks for your insightful feedback. Indeed, next-item prediction is a special case of UBTP when the prediction length k = 1, and it corresponds to the widely studied task of Sequential Recommendation (SR). UBTP requires the model to generate a coherent sequence of future interactions, capturing inter-item dependencies and preference evolution over multiple steps, which are not addressed in standard next-item recommendation settings.

To address the reviewer’s concern, we report the performance on the first position of the predicted trajectory (i.e., the first item among the multi-step prediction) using models trained under the UBTP setting. This provides a reference point that overlaps with next-item prediction. We report the performance on the first position (i.e., the next item) from the predicted trajectory. The results on MovieLens-1M (len=5), evaluated by HR@5 and NDCG@5, are shown below:

LPDO achieves the highest performance on the first predicted item, suggesting that even when evaluated on short-horizon prediction, the model trained for UBTP generalizes well. We will include these results in the appendix of the revised version.

---

W4&Q4: As this paper focuses on the UBTP task, the authors should also provide stronger baselines specifically designed for UBTP.

Thank you for your insightful suggestions. We would like to clarify that in our experiments, we have already extended SASRec to support UBTP by applying the Top-1 softmax at each prediction step (as SASRec-ar), enabling multi-step recommendation. This approach aligns with the UBTP setting and provides a fair baseline for comparison. Besides, we include sequential recommendation baselines such as DiffuRec and DreamRec, which have been widely used and offer valuable points of comparison. These baselines help demonstrate the effectiveness of our method.

Regarding the embedding size, our implementation does not require changing the embedding dimension from d to d×k. Instead, the search space expands from d to $d^k$, where d is the number of candidate items and k is the prediction length. This design ensures that the model can generate listwise predictions without modifying the underlying embedding structure.

Regarding incorporating the Plackett-Luce optimization ($\gamma=0.5$) with SASRec, we report the performance comparison below:

This experiment shows that incorporating PL optimization brings performance improvements to existing models such as SASRec. However, LPDO still outperforms these variants by a clear margin across all metrics, demonstrating that our method benefits not only from listwise supervision, but also from its integration into the diffusion-based trajectory modeling framework.

---

W5&Q5: Discussion on Existing Methods. The authors do not provide a comprehensive comparison with existing methods. I recommend the authors to compare LPDO with existing autoregressive recommenders (e.g., SASRec) and diffusion-based recommenders on the UBTP task (with K=1 and K > 1).

Thank you for raising this point. In our experiments, we include SASRec with autoregressive mode (denoted as SASRec-ar), as well as diffusion-based baselines such as DiffuRec and DCRec. We provide the comparion results of K=1 in the response of W3&Q3, please refer to it. As for results of K=3/5/10, please refer to Tab.1.

We appreciate your feedback. If there are additional aspects you would like us to elaborate on, we are happy to provide further analysis.

---

Q6: Typo at Line 237 and 466

Thank you for pointing out. We will revise the manuscript to correct this.

[1] Limitations of Autoregressive Models and Their Alternatives

[2] Fast and accurate inference of Plackett–Luce models

[3] Plackett-Luce model for the learning-to-rank task

[4] RankDistil: Knowledge Distillation for Ranking

[5] Self-Attentive Sequential Recommendation

[6] DiffuRec: A Diffusion Model for Sequential Recommendation

[7] Generate What You Prefer: Reshaping Sequential Recommendation via Guided Diffusion

Thanks for the authors' detailed rebuttal. While the rebuttal partially resolved my concerns, I still have some questions.

The definition of ListPref is quite confusing. The ListPref loss (11), or the Soft-ListMLE loss (9), needs the full ground-truth ranking label for the $K$ candidates $A _ {u, j}$ at time step $j$. However, in sequential recommendation, there is only one positive item $i _ {u, n + j}$ at each time step $j$. Therefore, the full ground-truth ranking label can only be one-hot, i.e., the positive item $i _ {u, n + j}$ should be ranked at the top, and the others can be ranked arbitrarily, no matter how many candidates $K$ are predicted. From this perspective, considering there is no need to capture the "relative ordering among multiple relevant items" (since there is only one relevant item), I question the authors' response to Q2(i).

In Q2(ii), the authors mentioned that "LPDO interacts with all candidate items for optimization without negative sampling". While I agree with this statement, I still concern that why ListPref (11) is independent of the number of candidates $K$ that predicted at each step? It is apparent that $K$ appears in the Eq. (11).

I may reconsider my rating if the authors can clarify the above issues, and I thank the authors again for their efforts in rebuttal.

Thank you for your valuable comments and thoughtful follow-up. We appreciate the opportunity to clarify the formulation of ListPref in the context of our task.

The definition of ListPref is quite confusing. The ListPref loss (11), or the Soft-ListMLE loss (9), needs the full ground-truth ranking label for the $K$ candidates $A_{u,j}$ at time step $j$. However, in sequential recommendation, there is only one positive item $i_{u,n+j}$ at each time step $j$. Therefore, the full ground-truth ranking label can only be one-hot, i.e., the positive item should be ranked at the top, and the others can be ranked arbitrarily, no matter how many candidates $K$ are predicted. From this perspective, considering there is no need to capture the ‘relative ordering among multiple relevant items’ (since there is only one relevant item), I question the authors’ response to Q2(i).

We understand the reviewer’s concern, and we believe it stems from a possible mismatch between two task formulations. We greatly appreciate the chance to further distinguish our setting from the conventional next-item recommendation task.

1. Step-by-step next-item recommendation

   • Definition: At each time step $j$, predict only the single next item $i_{u,n+j}$.
   • Ground-truth label: One-hot (one positive, all others negative).
   • Implication: Listwise losses over a one-item list reduce to a Top-1 choice model.

2. Whole-sequence prediction (our setting)

   • Definition: Predict the entire length-$K$ future trajectory $\bigl(i_{u,n+1},i_{u,n+2},\dots,i_{u,n+K}\bigr)$ as one ordered list.
   • Ground-truth label: A full permutation of $K$ distinct positive items—not a single one-hot.
   • Implication: ListPref (Eq.(11)) directly maximizes the Plackett–Luce likelihood of this complete sequence, requiring and utilizing a full ranking of $K$ positives.

Why the one-hot critique does not apply in our setting for UBTP task.

• Full-sequence ground truth: In our UBTP task, the target is the ordered sequence $\bigl(i_{u,n+1},\dots, i_{u,n+K}\bigr)$. We therefore have exactly one positive per rank, for a total of $K$ positives, which constitutes a full ranking label.
• ListPref’s mechanism: ListPref still performs sequential softmax draws over all $K$ candidates, where each position’s softmax is conditioned on all previous positions. This setting enable rich comparisons at each draw (positive vs. negatives and negative vs. negative), but now over the complete sequence rather than a single next-item step.
• No one-hot reduction: Unlike stepwise training, our formulation never reduces to “one positive vs. all negatives” in isolation; it always conditions on previously drawn items, fully leveraging the permutation.

Empirical validation.

In Sec.5.2 (Tab.1), LPDO’s ListPref loss outperforms the position-wise CE baseline (SASRec-ar). To further validate this, we provide a performance comparison on the first position of the prediction sequence on MoiveLens-1M (target len=5) as below. Our LPDO also surpasses SASRec-ar in this setting. These gains confirm that, under whole-sequence prediction, ListPref harnesses the full permutation of $K$ positives to deliver superior ranking quality.

In addition to empirically validating that our proposed loss achieves higher HR@5 and NDCG@5 compared to position-wise CE loss, we would also like to highlight our contributions beyond loss formulation. While the ListPref loss is a component of our design, the key contribution of LPDO also includes integrating listwise supervision within a non-autoregressive diffusion framework for multi-step trajectory generation. This enables:

• Coherent prediction across multiple future steps.
• Jointly optimized reconstruction + ranking via a principled ELBO.
• Causal modeling via transformer backbones. To our knowledge, this is the first work to bridge listwise learning and diffusion-based trajectory prediction, moving beyond conventional autoregressive or pointwise paradigms.

In Q2(ii), the authors mentioned that "LPDO interacts with all candidate items for optimization without negative sampling". While I agree with this statement, I still concern that why ListPref (11) is independent of the number of candidates predicted at each step? It is apparent that K appears in the Eq. (11).

Thank you for the suggestion related to our presentation of the $K$.

In the standard sequential-recommendation setting (e.g., DiffuRec[1], DreamRec[2]), they adopt all items in the dataset for both training and evaluation. We follow this setting in our paper, and the notation $K$ refers to using the entire item set (i.e., $K$ = all items) as candidates for optimizing the predicted items.

However, this $K$ is different from the @K used in evaluation metrics (e.g., SNDCG@K, SeqMatch@K). In these metrics, @K denotes the number of top-ranked items considered when evaluating whether the ground-truth item (i.e., the item the user actually interacted with) appears in the recommended list. In the UBTP task, each predicted results in the sequence has its own top@K items.

We sincerely appreciate the reviewer for pointing this out, and we will revise the notation from @K to @N in the updated manuscript to avoid confusion and improve clarity.

[1] DiffuRec: A Diffusion Model for Sequential Recommendation

[2] Generate What You Prefer: Reshaping Sequential Recommendation via Guided Diffusion

We greatly appreciate your time and effort in reviewing our work. Your comments were very helpful, and we have provided detailed responses below to clarify and address the issues you raised.

W1: From Figure 3 (b), it is shown that the prediction results for long-term goals are similar to those for short-term goals, and LPDO shows a similar decrease in this prediction result compared to other sequence recommendation methods (pos1-pos5). This does not seem to reflect the superiority of LPDO in modeling users' long-term preferences compared to other methods.

Thank you for the insightful comments. We try to address these concerns from two distinct perspectives:

• Performance decay is inherent in the UBTP task. Compared to one-step prediction, UBTP is intrinsically more challenging due to error accumulation over multiple steps. As the model predicts further into the future, uncertainty compounds, and performance naturally declines due to error propagation. This is a well-known phenomenon in multi-step sequence modeling, such as the NLP task [1]. Despite this, LPDO demonstrates greater robustness and slower degradation than other methods, underscoring its effectiveness.

• A similar performance trend does not equate to having similar modeling capabilities. While it is true that all models, including LPDO, demonstrate a decreasing accuracy trend across future positions, this observation alone does not imply comparable modeling capacity. As illustrated in Fig.3(b), LPDO consistently outperforms all baselines at each prediction step, including the later stages. This consistent advantage highlights LPDO’s superior ability to capture both short-term signals and long-term user preferences.

We thank the reviewer for the insightful comment and acknowledge that there is still potential for enhancing the UBTP task in future work.

---

W2: The expression in the paper is not clear enough, such as the "UBTM" in line 149.

Thank you for pointing this out. ‘UBTM’ is a typo and should be ‘UBTP’. We will revise the manuscript to correct this.

---

W3: LPDO did not show a significant increase in advantages compared to other models at longer lengths (len=10) and shorter lengths (len=5).

Thank you for bringing this point to us. We would like to clarify that LPDO consistently demonstrates substantial performance improvements across both shorter (len=5) and longer (len=10) future sequence prediction as shown in Tab.1.

For instance, on the Last-FM (target len=5) setting, LPDO achieves a 29.83% improvement over the best baseline in SeqHR@5, and a 36.81% absolute gain in SeqMatch@50, which measures strict trajectory-level accuracy. On the LastFM (target len=10) dataset, LPDO improves SeqHR@5 by 30.05% and SeqMatch@50 by 6.87% over the best baseline. These gains remain substantial and statistically significant, and they underscore LPDO’s robustness across different settings.

---

W4: Performance heavily depends on manual tuning of loss weight $\lambda$ and penalty $\gamma$.

Thank you for bringing this point to us. We would like to clarify that loss weight $\lambda$ can be set to 0.1 for all experiment settings to achieve superior performance, as mentioned in Sec.5.1. As for the penalty factor $\gamma$, as shown in Fig.3(c) and the additional results below, all different $\gamma$ settings consistently outperform the baselines in Tab.1, demonstrating the robustness of our method to this hyperparameter.

Table: MovieLens-1M (len=5)

Table: LastFM (len=5)

We will add additional results in the revised version to improve the clarity of $\gamma$ settings. Thank you for your suggestion regarding adaptive tuning. We will consider incorporating adaptive strategies in future work.

---

W5: The model requires fixed-length trajectories during training. Real-world sequences are variable-length, limiting applicability to open-ended scenarios.

Thank you for raising this concern. We would like to clarify that our model can support variable-length sequences during training. Our model is designed for the UBTP task (next-sequence prediction), which is an extension of the next-item prediction setting in sequential recommendation. In this field, it is standard practice to use sequences with a maximum length for training, and variable-length sequences are typically handled by padding shorter sequences to the maximum length, which is a widely adopted and effective strategy in sequence modeling. Therefore, our method can naturally support variable-length trajectories without loss of generality or applicability. During model inference, our model can also support variable-length sequences.

Thus, our method remains applicable to open-ended scenarios and real-world deployment. We will clarify this implementation detail in the revised version.

---

W6: Why is it that setting the diffusion step size to 50 and the sampling step size to 1 can yield very significant results?

Thank you for your question. Compared to CV [2] and NLP [3] tasks, recommendation targets lie in a low-dimensional, discrete, and highly structured space (e.g., item indices). The conditional signals, such as user history, are strong and tightly constrain the output manifold. As a result, even coarse denoising with a single sampling step can reliably recover high-quality item representations for ranking. In contrast to NLP/CV tasks requiring fine-grained generation, the discrete and sparse nature of recommender outputs makes them significantly more robust to approximation during fast sampling. Meanwhile, our observation is consistent with prior works such as DiffRec[4], SdifRec[5], and DCRec[6], which also report that a small number of steps can yield competitive results in recommendation tasks. We will clarify this point in the revised version.

[1] Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks

[2] High-Resolution Image Synthesis with Latent Diffusion Models

[3] DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models

[4] Diffusion Recommender Model

[5] Bridging User Dynamics: Transforming Sequential Recommendations with Schrödinger Bridge and Diffusion Models

[6] Dual Conditional Diffusion Models for Sequential Recommendation

We are truly grateful for the time and effort you dedicated to reviewing our work. We hope our detailed responses have clarified the issues raised.

W1&Q1: The key innovation and insights are a bit unclear. Lines 53–59 are too technical for the introduction. Fig.1(c) is not echoed by the technical presentation.

Thank you for your thoughtful feedback and suggestions. The key advantage of our proposed LPDO method is that it directly optimizes the listwise preference of item sequences as shown in Fig.1(d), rather than optimizing individual items independently as in pointwise methods. This listwise optimization allows LPDO to better capture the dependencies and overall ranking quality of the entire sequence, leading to improved recommendation performance. Figure 1(c) is meant to conceptually illustrate how our preference-aware diffusion model focuses the prediction distribution more effectively on semantically related targets. We elaborate this effect formally in Section 4.2 (Bridging Diffusion and Sequential Listwise Ranking) and Section 4.4 (Connecting Diffusion Models With Listwise Maximum Likelihood Estimation), where we introduce the ListPref loss and the ELBO formulation. By aligning the diffusion denoising steps with the Plackett–Luce listwise supervision signal, the model is guided to concentrate generation probability on ranked, preference-consistent item sequences, thereby reducing the dispersion toward unrelated content. We will clarify this in revised version.

---

W2: The empirical analysis is a bit unclear about the key contributor to the improved performance of the proposed method. There could be a couple of factors why the proposed method outperformed previous methods: (1) k-step prediction; (2) novel loss function; (3) causal transformer.

Thank you for raising this point. As you mentioned, the performance improvement of LPDO mainly comes from three factors:

• Designed for k-step prediction: LPDO is specifically designed for the k-step prediction task. While all baselines are extended to support prediction over k future time steps for fair comparison, LPDO is the first model explicitly designed for the UBTP task, addressing the limitations of directly applying next-item prediction models in this setting.

• The novel loss function: We proposed a novel listwise preference-aware loss. Our loss maximizes the joint likelihood of the entire ordered sequence, better capturing the ordering and dependencies among items. We have provided an ablation study of different loss components in LPDO in Tab.2. The results demonstrate the effectiveness of our loss. Furthermore, we replace the original list preference loss in LPDO with traditional cross-entropy loss for directly comparison. The results on ML-1M (target length=5) presented below indicate the importance of list-wise optimization in the UBTP task.

• Applying causal transformer: Regarding transformer modules, we provide a comparison between different transformer modules, including bidirectional [1], prefix [2], causal [3], on ML-1M (target length=5) as below. The results show that causal transformer outperforms the other two types, which indicates the importance of capturing performance and causality among user iterations.

---

W3: Since the proposed method predicts a longer future, would it require more data or a longer time to train?

Thank you for bringing up this point. It is worth noting that both LPDO and other baselines demand more training data to reach optimal performance. This is largely because the UBTP task entails a significantly larger solution space than next-item prediction, expanding from $n$ to $n^k$, where $n$ is the number of candidate items and $k$ is the number of steps to predict.

To further investigate this issue, we analyze model performance (measured by SeqMatch@50) on the ML-1M dataset (target length = 5) under varying training set sizes: the full dataset, 3,000 samples, and 1,000 samples. As shown below, model performance improves consistently with more training data. Notably, LPDO consistently outperforms all baselines across all three data settings.

Regarding training time, as shown in Fig. 3(a), our model converges faster than the baseline due to LPDO`s optimization for global sequence-level preference. Thus, LPDO does not increase the training cost compared to other baselines; rather, it enhances both computational efficiency and predictive effectiveness.

---

W4: Line 119: "target item" is confusing.

Thank you for your insightful feedback. The subsection around Line 119 is intended to introduce diffusion models and their general application in recommender systems. The specific application of diffusion to the UBTP task, which involves predicting a sequence of k items, is detailed in Section 4.1. We will revise the manuscript to clarify this distinction and avoid potential confusion.

---

W5&W6: Line 158 and Line 198 is mistakenly referred.

Thank you for pointing out the incorrect references. We will carefully review and correct these references in the revised version.

---

Q2: Line 153 mentioned that the paper follows [18] to also include history H as part of X0, which the model is trying to generate.

Thank you for this insightful question about our design choice! The concatenation of history H and target z is first proposed by DiffuSeq [4] in the NLP task, and then DCRec [5] introduce it in the sequential recommendation task for conditioning diffusion models. This can better capture detailed sequential patterns and contextual information from the user’s history [5]. Based on this, we introduce causal attention to further capture the preference evolution. As shown in Fig.2, the reverse/generation process starts from $X_T$, where $H_T$ and $z_T$ are noised. $H_t$ in each diffusion steps can provide extra historical user information during the generation until obtaining clean $H_0$ and $Z_0$.

Empirically, we observe that this concatenation-based conditioning significantly improves the guidance signal during generation compared to approaches that do not explicitly model the historical context in this manner (e.g., DreamRec [6], PreferDiff [7]). As a result, our method achieves more accurate and user-preferred recommendations.

---

Q3: For each $t$, the history $H$ is already a part of $X_t$. Why do we have the extra $H$ in $p_\theta(X_{t-1}| X_t, H)$?

Thank you for the insightful comment. We would like to clarify that, in addition to the $H$ component within $X_t$, there is an extra $H$ used as direct conditioning, as illustrated in Fig.2. This additional conditioning serves as the input to the cross-attention layer and plays a crucial role in guiding the generation process. This strategy was first introduced in Stable Diffusion [8] and later adopted in DCRec [5] for recommendation tasks. We thank the reviewer for this insightful question and will enhance the explanation accordingly in the revised manuscript.

---

Q4: It seems that $L_{ListPref}$ is defined for a particular future time step j? How would the system combine all the k future time steps in the loss?

Thank you for highlighting the potential ambiguity in our presentation of the listwise loss. Equation (11) was framed as “per‐step” $L_{\rm ListPref}(j)$ simply to keep the expression compact and to focus the reader’s attention on the mechanics of the listwise term at a single horizon $j$. In fact, the full listwise objective used in all our experiments is the sum of these per‐step losses over $j=1,\ldots,k$, exactly as shown (and implemented) in the model code and in Equation (16). By summing across all $k$ positions and employing our cross‑step attention blocks inside the Cross‑Conditional Diffusion Transformer, the model naturally receives joint gradient signals that enforce causal consistency between early and later predictions. In this way, the loss truly operates on the entire future sequence, not just on each step in isolation.

---

Q5: Eq.16 contains both reconstruction loss and list preference loss. I thought the reconstruction loss already contains the list preference loss.

Thank you for raising this point. We would like to clarify that Eq.(16) is derived from Eq.(5), with the primary modification being the introduction of the list preference loss, a novel ranking objective specifically designed for the UBTP task, replacing the original ranking loss in Eq.(5). Meanwhile, the reconstruction loss and regularization loss are preserved. We appreciate this constructive feedback and will revise the manuscript accordingly to enhance clarity.

[1] BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

[2] Exploring the limits of transfer learning with a unified text-to-text transformer

[3] Language Models are Unsupervised Multitask Learners

[4] Diffuseq: Sequence to sequence text generation with diffusion models

[5] Dual conditional diffusion models for sequential recommendation

[6] Generate What You Prefer: Reshaping Sequential Recommendation via Guided Diffusion

[7] Preference Diffusion for Recommendation

[8] High-Resolution Image Synthesis with Latent Diffusion Models

We sincerely thank the reviewer for the detailed, thoughtful, and positive evaluation of our work. We are especially grateful for your recognition of our contributions. Your comments on the clarity of presentation, strength of empirical results, and potential impact are highly encouraging and greatly appreciated.

W2, Q1: The lack of open-source code.

Thank you for raising this point. We plan to release the data and code after the paper is published.

---

W3, Q2: Some further clarity about which of the experimental results are statistically significant would be appreciated.

Thank you for the suggestion. We confirm that the improvements are statistically significant for the main performance results (Sec. 5.1), factor analysis (Sec. 5.2), and ablation study (Sec. 5.3). We will clarify this in the revised manuscript and ensure that statistical significance is reported for all other experiments (in appendix).