We sincerely appreciate the reviewer's valuable time and insightful comments. Below are our point-by-point responses to the questions raised. We look forward to any additional feedback the reviewer may provide.

Response to W1 & Q2

We sincerely thank the reviewer for recognizing the high inference efficiency of our method. In the main text, we have discussed inference efficiency, and we have included the training and inference algorithm workflows in Algorithm 1 and Algorithm 2 in the appendix. Here, we provide a detailed explanation of the inference and training costs.

• Inference Efficiency: During inference, only the recommendation backbone is utilized. The contrastive learning tasks and the LLMs’ semantic embeddings are not involved in the inference process. This ensures that our framework can be deployed in real-world applications without incurring any additional inference latency from incorporating LLMs.
• Training Efficiency & Cost: The training process of our method consists of two stages: In the first stage, we use an LLM API to obtain semantic information and convert it into embeddings, which are then cached to construct contrastive sample indices. The primary time cost in this stage comes from the API calls. However, by employing asynchronous concurrency, this step can be completed within a few hours. Crucially, this stage is performed once and requires no repetition during model training. Regarding the monetary cost of API usage, if using the DeepSeek API, all the experiments in this work can be covered within a budget of $50. Alternatively, if using an open-source model like Qwen-32B, this cost is reduced to zero, as it can be run locally without incurring API fees. In the second stage, we use the pre-constructed contrastive sample index to train the recommendation model. Regarding the computational complexity of this stage, our method maintains comparable time complexity to general ID-based contrastive recommendation approaches. The only additional overhead during training compared to conventional contrastive recommendation models comes from the learnable sample synthesis module—a single attention layer—whose parameter size is negligible compared to that of the main recommendation model.

Response to W2 & Q1

• We sincerely apologize for any confusion caused. In fact, we have already addressed this question in the original paper. Specifically, Lines 202–208 on Page 6 clarify why the Learnable Sample Synthesis is applied only to user-level contrastive learning and not extended to the item level. Due to space constraints, we provided further elaboration in Appendix C.1, along with experimental comparisons in Table 4 of the appendix.
• To reiterate, our preliminary experiments explored employing learnable synthesizers at the item level (analogous to inter-user contrastive learning) to generate substitute items. However, this approach yielded no measurable performance improvements (see Table 4, Appendix). We attribute this outcome to the inherent nature of item semantics, which are more straightforward to interpret and quantify than user preferences. Consequently, directly retrieving substitutes from semantically similar candidate pools proves both simpler and more reliable than attempting to match users with analogous preference patterns.

Response to Q3

We appreciate the insightful question regarding our choice of cosine similarity for semantic retrieval. Our decision was based on the following considerations:

• Normalization robustness: Unlike the dot product, cosine similarity inherently normalizes similarity scores by accounting for vector magnitudes. This property makes it more robust to scale inconsistencies in semantic embeddings, particularly since we do not apply explicit normalization to our embeddings.
• Alignment with the text embedding model’s pretraining objective: The text embedding model we employed, SimCSE-RoBERTa, was explicitly trained using cosine similarity as its semantic similarity metric. By adopting the same measure for retrieval, we ensure consistency with the model’s original training objective, thereby enhancing the reliability of user similarity computations.

We sincerely appreciate the reviewer's valuable time and insightful comments. Below are our point-by-point responses to the questions raised. We look forward to any additional feedback the reviewer may provide.

Response to W1: Justification of Motivation

We sincerely apologize for any misunderstanding, and we appreciate this opportunity to clarify our motivation.

• As presented in Page 2, lines 41–44, we posit: "Many existing methods construct contrastive pairs through random augmentation operations such as random masking and Dropout. However, in sequential recommendation where data is inherently sparse and exhibits sequential patterns, such random operations may lead to a complete change in the sequence’s semantics." Specifically, our critique focuses on random data augmentation methods, wherein contrastive samples are constructed in an entirely arbitrary fashion, frequently compromising the semantic of the original sequence. In contrast, our method is based on LLM's semantic comprehension. It stands to reason that our approach would yield more reliable results. This claim has been empirically validated through both ablation studies (Table 2 "w/o semantic") and comparative experiments (Table 1).

• To further mitigate the potential gap between LLM-summarized user preferences and genuine user interests, we avoid naively adopting LLM-identified "most relevant" users as positive contrastive samples. Instead, we propose a Learnable Sample Synthesis approach, where the LLM only performs initial screening to generate a candidate set of potential positive samples, and then a trainable synthesis mechanism is used to dynamically generate informative samples from the candidate set, which is optimized end-to-end based on the final recommendation performance. This ensures the synthesized positive samples enhance recommendation metrics.

• Explanation of Figure 4. The figure illustrates the impact of k on model performance. When k=1, the method simply adopts the top-1 most relevant sample selected by the LLM as the contrastive sample, bypassing our learnable synthesis module. The suboptimal performance at this setting indicates a potential gap between the LLM's inferred user preferences and actual user preferences, underscoring that leveraging LLMs to guide contrastive sample construction is indeed nontrivial. As k increases (e.g., to k=10), the LLM performs preliminary screening to select k potentially relevant candidates for the sample synthesis module. Through end-to-end learning optimization, our module synthesizes an optimal contrastive sample from these candidates. The performance improvement indicates that the learnable synthesizer effectively mitigates the semantic gap between LLM outputs and real user preferences, thereby generating superior contrastive samples to improve the performance. Excessively large k values (k=100) relax the LLM's semantic similarity threshold, introducing substantial noise, which compromises the learning process and degrades performance. Overall, this figure highlights our key contribution: developing an effective learnable approach to bridge the semantic gap between LLM outputs and genuine user preferences.

Response to W2 & Q1: Experimental Validation of the Learnable Sample Synthesizer

1) Clarification of Motivations

• The statement, "User preferences exhibit significant heterogeneity across individuals," highlights the diversity of user interests. The primary goal of recommendation models is to effectively model this diverse preferences. However, existing methods construct contrastive pairs through random augmentation or predefined hard rules, which significantly alter the user preferences, potentially leading to substantial differences in preferences between two users treated as positive samples. Pulling users with differing interests closer in the latent space leads to representation homogenization, which contradicts the core objective of recommendation.

• Our method leverages LLM-based semantic information to perform an initial screening and generate a candidate set that aligns with each user's preference-related semantics, which helps alleviate the aforementioned issue. However, directly using the top-1 sample as the positive sample for contrastive learning may lead to a potential problem: “There may exist a gap between the LLM’s interpretation of preferences and the user’s actual preferences.” To address this issue, we propose the Learnable Contrastive Sample Synthesis module.

2) Experimental Validation

In the ablation study section of our original paper (Table 2), we reported the experimental results for the "w/o learn." setting. The observed performance degradation validates the effectiveness of the proposed Learnable Contrastive Sample Synthesis module.

We greatly appreciate the reviewer’s suggestion. In response, we conducted two additional experiments to validate the motivation, which will be added to our final version.

• To validate whether there is a gap between the LLM's understanding of user preferences and the actual user preferences, we randomly selected 100 users and, for each user, observed and compared the semantic similarity score distribution of the users in the candidate set (as derived from LLM-based relevance) with the weight distribution learned by the proposed learnable module for these users. The results show that the ranking results differ in 99% of the cases. Due to space limitations, we provide an example of the selection probabilities comparison as follows:

From the results, we can observe that there are clear differences between the two distributions, which validates the rationale behind employing the proposed learnable module. Our learnable module leverages a non-linear adapter and an attention mechanism to automatically assign weights to the samples in the LLM-filtered candidate set, which makes that the selection probabilities of candidate samples are no longer solely based on the original semantic embeddings but are instead learned and optimized end-to-end according to the objective of the recommendation model.

• To validate the impact of the learnable module on representation heterogenization, we visualized the user sequence representations learned by our method with and without the learnable module using t-SNE. Due to rebuttal guidelines prohibiting the inclusion of images or links, we provide a brief summary of the results here. The embeddings with the learnable module are more dispersed, indicating greater heterogeneity in the representations, while the embeddings without the module are relatively compact. The results suggest that the learnable module effectively mitigates the risk of user representation homogenization or even collapse, preserving the heterogeneous diversity of user interests.

Response to W3 & Q2: SOTA LLM-based Text Embedding Models

We sincerely appreciate the reviewer's suggestions regarding the text embedding model. We selected sup-simcse-roberta-large for two reasons: 1) It is a widely adopted model for extracting sentence-level semantic embeddings, and 2) Since this component is not our primary innovation focus, using a general embedding model better demonstrates that our method's performance does not rely on advanced text embedding models.

We fully agree that comparing LLM-based text embedding models would yield valuable insights. In response, we have extended our evaluation to incorporate two additional SOTA embedding models: 1) OpenAI’s text-ada-embedding-002, and 2) Qwen3-Embedding-8B. The results are shown as the following table:

Our results show that advanced text embedding models can boost the performance of our method: 1) text-ada-embedding-002 (1B params, 1536D) outperforms SimCSE-RoBERTa-large (335M params, 1024D) across all datasets, thanks to its greater capacity for semantic representation. 2) Qwen3-Embedding-8B (8B params, 4096D) achieves SOTA results, surpassing both models in preference modeling accuracy.

Furthermore, to demonstrate that our method consistently outperforms other LLM-based approaches across different text embedding models, we compared our method with the strongest LLM baseline, RLMRec, under the best-performing Qwen3-Embedding-8B.

Dear Reviewer UspK,

We sincerely appreciate the time and effort you've dedicated to reviewing our work. We have carefully addressed all your comments in our rebuttal, including detailed explanations and experiments regarding:

• The experimental validation of our motivation, and

• The SOTA LLM-based embedding models.

Please let us know if these responses have adequately addressed your concerns.

Thank you again for your valuable insights. We look forward to your reply.

We sincerely appreciate the reviewer's valuable time and insightful comments. Below are our point-by-point responses to the questions raised. We look forward to any additional feedback the reviewer may provide.

Response to Weakness 1

We sincerely appreciate the reviewer’s valuable suggestion. Below, we answer this question in two aspects:

• Our method does not rely on LLM inference during the prediction phase; instead, it only involves the inference of the backbone recommendation model. Therefore, our framework can be deployed in real-world large-scale applications without introducing additional latency from LLM integration. This also explains why our initial experiments did not explore smaller LLMs.
• We fully acknowledge the reviewer’s insightful suggestion regarding smaller LLMs, as they could reduce training time and computational costs. Accordingly, we have conducted additional experiments using Llama3.2-3B-Instruct, and the results are presented below.

The experimental results indicate that the 3B LLM underperforms compared to larger models. We attribute this to the fact that reasoning and summarizing user preferences require advanced inferential and abstraction capabilities, which smaller models tend to exhibit to a lesser degree than their larger counterparts. Consequently, the quality of generated user preference descriptions may be comparatively lower. In contrast, Qwen-32B, despite having fewer parameters than DeepSeek-V3, still maintains a sufficiently large scale to preserve most of these capabilities, thus showing less significant performance degradation. These observations suggest a potential trade-off between model scale and performance - smaller models may offer improved training efficiency, but correlate with diminished effectiveness in downstream tasks. We will include the experimental results and analysis in the appendix of the final version.

Response to Weakness 2

• In fact, we had already conducted the experiment mentioned by the reviewer, and the results are presented in Table 2 of the Ablation Study section in the original paper. The "w/o LLM" setting denotes the scenario where the LLM is not used for text comprehension and processing—that is, the SimCSE-RoBERTa model directly encodes user and item texts, as the reviewer proposed.
• The results in Table 2 indicates directly applying the text embedding model without LLM-based processing leads to performance degradation. This confirms that leveraging the LLM’s ability to understand and reason about user preferences is crucial. Without secondary processing by the LLM, the textual data from user historical behaviors becomes excessively verbose, with key information obscured by irrelevant details. Such lengthy and noisy text poses significant challenges for the embedding model, resulting in lower-quality embeddings. The LLM’s reasoning and summarization capabilities effectively mitigate this issue.

Response to Weakness 3

We sincerely thank the reviewer for their insightful question and suggestion. The primary application scenario of this study focuses on cases where textual attributes of items are available, consistent with prior work [1, 2]. The metadata in the datasets used in our experiments exhibits a certain degree of sparsity, typically below 20%. We noticed that most real-world applications, such as Amazon, Taobao, Netflix and YouTube, generally provide and maintain high-quality textual metadata. For scenarios where text metadata is extremely sparse or entirely unavailable, such cases pose significant challenges for all LLM-based recommendation methods. However, addressing these challenges is beyond the scope of this paper due to space and time constraints. We plan to explore this direction further in future work. Once again, we deeply appreciate the reviewer's valuable suggestion.

[1] Ren, Xubin, et al. Representation learning with large language models for recommendation. WWW' 24

[2] Liu, Qidong, et al. Llm-esr: Large language models enhancement for long-tailed sequential recommendation. NeurIPS' 24

Response to Weakness 4

We sincerely thank the reviewer for their suggestion regarding the Conclusion section. Your feedback will help improve our paper further. In the final version, we will expand the Conclusion and include a discussion on future work. The revised Conclusion is as follows:

"In this paper, we analyze the limitations of contrastive learning in sequential recommendation, namely Semantic Divergence and Unlearnability. To address these issues, we propose SRA-CL, a novel framework that enhances contrastive sample construction by integrating LLM-based semantic retrieval with a learnable sample synthesizer. SRA-CL leverages the capabilities of LLMs without increasing the inference time of the recommendation model, making it practical for large-scale real-world applications. Through comprehensive experiments, we demonstrate that LLM-based semantic-guided contrastive sample construction improves the contrastive learning, and we validate the effectiveness of the learnable sample synthesis mechanism. Furthermore, experiments with different recommendation model backbones confirm the model-agnostic nature of our approach. Looking forward, there are potential directions for future work: (1) optimizing the computational overhead of LLMs during the training process, and (2) exploring how to adapt our method to scenarios with scarce textual information."

We sincerely appreciate the reviewer's valuable time and insightful comments. Below are our point-by-point responses to the questions raised. We look forward to any additional feedback the reviewer may provide.

Response to W1: About the Contribution and Positioning

1) The Contribution and Significance:

• Our work provides an in-depth analysis and novel exploration of contrastive learning in recommender systems. Unlike CV or NLP, contrastive learning in recommender systems presents unique challenges that render direct adoption of methods from other domains ineffective.
• Furthermore, while existing contrastive learning approaches in sequential recommendation have predominantly relied on ID-based representations, we critically examine their limitations and, for the first time, propose leveraging semantic information to construct higher-quality learnable contrastive pairs, which is achieved by harnessing the understanding capabilities of LLMs.
• We believe our work offers a significant contribution by redefining the understanding and application of contrastive learning in recommender systems. Our approach represents a meaningful and successful innovation, as mentioned by Reviewer 6Nm1: "The innovative use of LLMs to uncover semantically rich representations offers a promising new direction for further research and real-world application."

2) On the Positioning:

• Our paper aligns with NeurIPS's scope. Regarding the comment of “this paper is better positioned at an IR conference rather than NeurIPS”, we respectfully clarify that while specialized domains have dedicated venues (e.g., CVPR for computer vision, ACL for NLP, SIGIR for IR), NeurIPS serves as a premier interdisciplinary platform that explicitly welcomes cross-domain contributions—including research bridging information retrieval and machine learning.
• Critically, our work focuses on user/item understanding and contrastive learning—precisely the areas highlighted by the reviewer as "NeurIPS's emphasis on understanding/learning."
• Moreover, the acceptance of 21 papers at NeurIPS 2024 featuring the terms "recommendation" or "recommender" in their titles further validates the suitability of our work for this venue. Representative examples include:
  • Q. Liu et al. "LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation"
  • Y. Liu et al. "End-to-End Learnable Clustering for Intent Learning in Recommendation"

Response to W2: About the Efficiency

We appreciate the reviewer’s attention to efficiency. Actually, we have clarified the inference efficiency in Section 3.3 (Lines 220–224 of Page 6).

• Inference Efficiency: During inference, only the lightweight backbone recommendation model is executed, while contrastive learning tasks and semantic embeddings are entirely disabled. This design ensures no additional computational overhead from LLM integration, preserving the same inference latency as the original recommendation backbone. Thus, our approach remains deployment-friendly and scalable for real-world systems, as it introduces no runtime cost to the production pipeline.
• Training Efficiency & Cost: The training process of our method consists of two stages: In the first stage, we use an LLM API to obtain semantic information and convert it into embeddings, which are then cached to construct contrastive sample indices. The primary time cost in this stage comes from the API calls. However, by employing asynchronous concurrency, this step can be completed within a few hours. Crucially, this stage is performed once and requires no repetition during model training. Regarding the monetary cost of API usage, if using the DeepSeek API, all the experiments in this work can be covered within a budget of $50. Alternatively, if using an open-source model like Qwen-32B, this cost is reduced to zero, as it can be run locally without incurring API fees. In the second stage, we use the pre-constructed contrastive sample index to train the recommendation model. Regarding the computational complexity of this stage, our method maintains comparable time complexity to general ID-based contrastive recommendation approaches. The only additional overhead during training compared to conventional contrastive recommendation models comes from the learnable sample synthesis module—an adapter and a single attention layer—whose parameter size is negligible compared to that of the main recommendation model. Specifically, each experiment was conducted on a single 32GB V100 GPU and completed within 3 hours of training.

Response to W3: Discussion of LLM-based Methods

Our paper has provided a comprehensive comparison and discussion of state-of-the-art LLM-based sequential recommendation methods. This is demonstrated through:

1. Experiments: Our experimental design includes three SOTA LLM-based sequential recommendation methods as baselines, all of which were recently published in top-tier conferences such as NeurIPS and SIGIR. Additionally, we have compared against seven contrastive learning approaches and three classical recommendation methods, resulting in a total of 13 baseline methods. We believe this constitutes a thorough comparative studies in this domain as the primary objective of our work is to enhance existing contrastive learning methods in sequential recommendation using LLMs.
2. Related Work: In the related work section (Section 5), we have systematically reviewed the latest advancements in the topic of LLMs for sequential recommendation and clearly delineated the distinctions between our work and these existing approaches.

Below is the summarization of the comparison in terms of performance, efficiency, cost, and deployment difficulty:

• Performance: Our approach demonstrates consistently better recommendation accuracy across multiple benchmarks, validating its effectiveness.
• Computational Efficiency and Cost: 1) Inference Efficiency: Our method eliminates the need for LLM involvement or semantic processing during inference, making it as efficient as traditional recommendation models. 2) Training Cost: While leveraging semantic information, our framework maintains comparable training costs to other LLM-based methods.
• Scalability & Deployability: Due to its lightweight inference mechanism, our approach is highly scalable and readily applicable to real-world large-scale recommendation systems.

Response to Reviewer's Comment Regarding Large-scale Data Applicability

Our method is deployment-friendly and scalable for real-world applications, as it introduces no additional inference latency from incorporating LLMs. As discussed in Line 220-224 of Page 6, during inference, only the recommendation backbone is utilized. The contrastive learning tasks and LLMs’ semantic embeddings are not involved in the inference process. This design ensures no additional computational overhead from LLM integration, preserving the same inference latency as the original recommendation backbone.

Dear Reviewer rXdN,

We sincerely appreciate the time and effort you've dedicated to reviewing our work. We have carefully addressed all your comments in our rebuttal regarding:

• Contribution and Positioning,
• Efficiency,
• Discussion of LLM-based Recommendation Methods

Please let us know if these responses have adequately addressed your concerns.

Thank you again for your valuable insights. We look forward to your reply.