Thank you for your time and effort. Some of your suggestions are highly valuable and worth adopting. However, there are a few misunderstandings regarding the contributions of our paper and its relevance to the conference topic. We have provided detailed explanations and conducted additional experiments to address each of your concerns, and we hope this helps to clarify any confusion!

W1&W2: Contribution

The work [1] provided us with significant inspiration, as it demonstrated the critical role of diffusion models in high-performance parameter reconstruction. However, in our work, the diffusion model serves as just one tool for parameter generation. It is necessary for us to restate our key contributions and innovations: (1) the definition of Controllable Multi-Task Recommendation (CMTR), (2) the "backbone + task-specific adapter" structure in recommendation models, and (3) the use of conditional training to achieve the "one instruction, one model" paradigm. We acknowledge that our paper does not include A/B testing results, due to the company's complex processes and confidentiality requirements. Nevertheless, we conducted tests using real clickstream data collected from July 24, 2024, to August 24, 2024, as presented in the paper (Table 1), which reflects a certain level of consistency with the online tests.

Additionally, similar methodologies have already been adopted in transfer learning, as shown in work [2] published at ICLR 2024. This supports the relevance of adapting new technologies to different domains, aligning with the conference topic. Furthermore, our motivation is strongly rooted in industrial scenarios where business objectives frequently change. Compared to retraining models upon receiving new business goals, PadiRec eliminates retraining costs and shortens the response time from receiving a new objective to conducting a new model, as shown in Table 2 of our paper.

[1] Neural network diffusion. arXiv preprint arXiv:2402.13144 (2024).

[2] Spatio-temporal few-shot learning via diffusive neural network generation. ICLR, 2024

W3: SOTA Backbone

Thanks for reminding me about the optimal backbone. It's important to clarify that PadiRec does not focus on improving the accuracy of sequential recommendation models. Instead, it provides a framework that can adapt to any downstream recommendation model, with the goal of enabling customized recommendation models based on requirements without the need for retraining. Nevertheless, we have supplemented our experiments with the SOTA backbone [1] and will include the relevant references and results in the main text. The experimental results are shown below. The trend figure with the backbone LRURec has been added in Appendix § A.7 SOTA Backbone (LRURec).

[1] Linear recurrent units for sequential recommendation. WSDM, 2024.

Thanks for your response.

The ICLR community allows, encourages, and supports applied research papers. For instance, ICLR 2025 explicitly lists topics including, but not limited to:

• Applications in audio, speech, robotics, neuroscience, biology, or any other field.

Examples of applied works include those focusing on recommendation systems and retrieval [1, 2, 3, 4, 5], as well as studies effectively integrating diffusion models with various subfields based on their unique characteristics [6, 7, 8, 9]. Additionally, other trending applied research directions, such as LLM agents [10, 11, 12], are also actively encouraged by ICLR.

It is crucial to emphasize that our paper focuses on controllable multi-task recommendation, specifically addressing instant responsiveness and control based on dynamic changes in recommendation objectives. Parameter diffusion aligns perfectly with this goal, enabling test-time model customization via the "one instruction, one model" paradigm. Furthermore, conducting only offline evaluations on widely recognized datasets for recommendation algorithms is a standard practice [1, 2, 4, 5]. Moreover, the third dataset used in this study—industrial data—is an online dataset collected from a real-world production environment, further validating the practical value of our method.

Therefore, both in terms of topic and evaluation, this paper is well-aligned with ICLR's scope and standards. If there are additional technical concerns, we will do our best to address them.

[1] Towards Unified Multi-Modal Personalization: Large Vision-Language Models for Generative Recommendation and Beyond. ICLR, 2024

[2] Federated Recommendation with Additive Personalization. ICLR, 2024

[3] Sentence-level Prompts Benefit Composed Image Retrieval. ICLR, 2024

[4] LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation. ICLR, 2023

[5] StableDR: Stabilized Doubly Robust Learning for Recommendation on Data Missing Not at Random. ICLR, 2023

[6] Spatio-Temporal Few-Shot Learning via Diffusive Neural Network Generation. ICLR, 2024

[7] DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation. ICLR, 2024

[8] Multi-Source Diffusion Models for Simultaneous Music Generation and Separation. ICLR, 2024

[9] Training-free Multi-objective Diffusion Model for 3D Molecule Generation. ICLR, 2024

[10] AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. ICLR, 2024

[11] ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate. ICLR, 2024

[12] Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents. ICLR,2024

Thank you for taking the time to review our paper, especially given your busy schedule. Your suggestions are highly detailed, and your perspectives are insightful! However, some parts of your feedback reflect misunderstandings of our work. We have provided detailed explanations for each of your concerns and hope to clarify any misconceptions.

W1: Evaluation of Movielens

Valuable suggestions! The paper you mentioned [1] by Sun et al. reveals an interesting phenomenon and highlights some limitations of the MovieLens dataset. However, it was published on (October 18, 2024) after the submission deadline for ICLR 2025 (October 1, 2024). Therefore, the insights from this work cannot be considered as a reference for our ICLR submission.

Additionally, many classic sequential recommendation papers [2, 3, 4, 5] have included experiments based on the MovieLens dataset. Among them, the methods [2, 3]serve as our backbones. To align with these studies, we also adopt the MovieLens dataset. Besides, this paper also presents experiments on the Amazon dataset and real-world industrial datasets, further demonstrating the generalizability of PadiRec. However, we will take your feedback regarding the MovieLens evaluation into account in our future work. Thanks again for your valuable suggestions!

[1] Our Model Achieves Excellent Performance on MovieLens: What Does It Mean? TOIS, 2024

[2] Self-Attentive Sequential Recommendation. ICDM, 2018

[3] TiSASRec: Time Interval Aware Self-Attention for Sequential Recommendation. WSDM, 2020

[4] BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer. CIKM, 2019

[5] Generative Sequential Recommendation with GPTRec. SIGIR, 2023

I have reviewed the authors' response, but I still have concerns. The problem like using MovieLens's for sequential recommendation - I strongly recommend consulting the creators of MovieLens to understand why it is not suitable for sequential recommendation tasks. I am not convinced by using the existing literature. The Recsys models were not properly evaluated in the much literature. More broadly, I do not see how this work significantly advances the field given its overall contribution. While I respect the editor's final decision, I stand by my original score because I do not think the paper's contribution is sufficient to be accepted in ICLR.

"First, we must reiterate that the article discussing the limitations of MovieLens [1] was published after our submission."

Even without this paper, I still believe that researchers in this field should know what kind of tasks MovieLens can do. It is a rating prediction dataset, where uses were paid to rate movies. Many users can rate over 10 movies in 1-2 minutes. Rating sequences should not be considered as user real watching behavior sequences. The sequential patterns you captured come from the original MovieLens exposure strategy.

The following comments only represent personal opinions, and you can ignore them if you disagree.

1）Sequential recommendation on offline datasets has inherent limitations. The patterns we observe are largely artifacts of the original recommendation algorithms rather than true user behavior, which rarely follows strict sequences. This task was usually called session-based recommendation in the past, which acknowledges that user preferences tend to remain stable within short time windows.

2）Further improvements in offline accuracy metrics offer diminishing returns for these sequential task. While I understand academic researchers' limited access to online datasets, the field has reached established in terms of offline performance metrics.

Overall, I don’t think this paper is important to the Recsys community.

Thank you for your time, effort, and valuable suggestions! We have provided detailed explanations and supplemented the experiments based on your suggestions. We hope this is helpful to you, and once again, we sincerely appreciate your effort in reviewing our paper despite your busy schedule.

W1: More Utilities:

Thank you for your valuable suggestions. We have added user group fairness as a controllable metric. Specifically, we use the NDCG GAP@10 between male and female groups as the metric (a smaller value indicates greater fairness in NDCG@10 between the two groups) to evaluate the impact of fairness weights on the metric under different settings.

Given the large number of possible weight combinations for the three objectives, we explored the impact of fairness through a controlled variable approach:

1. Investigating the impact of fairness weight on other metrics (NDCG@10 and a-NDCG@10).

2. Examining the effect of fairness weight on its metric (NDCG-GAP@10).

The experimental results are presented below. The table records the performance of three metrics—accuracy (NDCG@10), diversity (a-NDCG@10), and fairness (NDCG-GAP@10)—under two conditions: unfair (fairness weight = 0.1) and fair (fairness weight = 1), as accuracy weight varies (constrained diversity weight = 1 - accuracy weight).

The line charts for the above metrics are presented in the PDF file, Appendix § A.9 More Objectives (Accuracy, Diversity, and Fairness), Figure 15. The following conclusions can be drawn:

1. Figures 15(a) and 15(b) show that the unfair and fair conditions have minimal impact on both NDCG@10 and a-NDCG@10 individually, as well as on the trade-off relationship between these two metrics.

2. Figure 15(c) demonstrates that the unfair and fair conditions have an impact on NDCG-GAP@10. Across multiple settings of accuracy weights, the NDCG-GAP@10 under the fair condition is consistently smaller than that under the unfair condition. This indicates that the control under the fair/unfair condition is effective.

To explore the fine-grained control of fairness weight, we investigated the performance of PadiRec on the three objectives under fairness weights of 0.1, 0.4, 0.7, and 1.0 when accuracy weight is set to 0.6 and 0.7 (at these point, both NDCG@10 and alpha-NDCG@10 demonstrate not bad performance). The results are shown in the table below (we have added it to Appendix § A.9 More Objectives (Accuracy, Diversity, and Fairness)):

Conclusion: As the fairness weight increases, accuracy (NDCG@10) and diversity (a-NDCG@10) show minimal fluctuation, while NDCG-GAP@10 steadily decreases, indicating improved fairness. This demonstrates that even under multiple objectives, PadiRec exhibits strong controllability.

Thank you for addressing my comments and concerns in your rebuttal. I appreciate the effort you put into clarifying my doubts and putting user group fairness into your metrics, which in my mind have made the paper more robust.

I want to emphasize that my initial score already reflected my positive view of the paper's quality and potential. While the rebuttal has resolved my concerns and improved the paper, there were no substantial changes that would warrant increasing the score further. As such, I have decided to maintain my original evaluation. Thank you for authors contributions to ICLR community.

Thank you very much for your recognition of our work! Here, we respond to each of your concerns regarding our study.

W1: Inference Complexity

As shown in Table 2, the diffusion model significantly reduces the time overhead from receiving a new task instruction to completing the model construction, compared to traditional retraining approaches.

Technically, our algorithm design has taken the inference cost into account from two aspects. One is the design to structure the recommendation model as a backbone + adapter architecture. The diffusion model is used solely to generate the adapter parameters, to reduce the training and generation overhead of diffusion. The other is the design of the denoising model of diffusion. We only stack 4 attention to act as a denoising model. The details of the diffusion model have been added to the PDF file and can be found in Appendix § A.13, Details of Diffusion. Additionally, we have added a comprehensive analysis of the computational cost, provided in Appendix § A.14 Diffusion Transformer FLOPs Calculation.

The conclusion of Appendix § A.14shows that our diffusion model requires only 0.9085 TFLOPs for the entire 500-step sampling process. Using the RTX 3090 as an example, which achieves 35.58 TFLOPS per second, the inference process for the diffusion model takes approximately 0.026 seconds. Including some data storage overhead, the total time remains within the order of seconds (as shown in Table 2, about 2.68s of SASRec). In real-world recommendation scenarios, it is generally acceptable for users to wait 2-3 seconds to customize a more personalized model.

W2: Joint Optimization

In §2 Problem Formulation and Analysis, we define controllable multi-task recommendation (CMTR) and distinguish it from multi-task recommendation (MTR). Under the definition of CMTR, a single task is defined as an optimization problem with a specific set of preference weights for multiple objectives. Therefore, a "task" under CMTR inherently involves the joint optimization of multiple objectives.