R$^2$ec: Towards Large Recommender Models with Reasoning

Thank you for your detailed and constructive review. We appreciate your recognition of our ablation study and empirical results. We acknowledge your concerns regarding inference latency, the limited user history setting, the absence of a pure pre-trained baseline, and the lack of cold-start analysis.

We will address each of these points in detail in our response below.

Q1: Inference Latency and Runtime Analysis

We appreciate the reviewer’s question regarding the runtime overhead introduced by the reasoning process.

First, we would like to clarify that the runtime of R²ec have been reported in Appendix G (“Analysis on Inference Efficiency”), the quantitative results on per-sample latency under different batch sizes.

Second, to provide more detailed comparison and analysis, we have expanded our latency evaluation with additional benchmarks using consistent hardware (RTX 3090) and unified implementation settings. Below, we present the runtime (in seconds per sample) for R²ec and its VLLM serving version, against various baselines.

These results demonstrate that R²ec achieves competitive efficiency among LLM-based recommenders, particularly when using modern inference frameworks. We attribute this efficiency in part to our use of scalable item embeddings, which avoids costly autoregressive token decoding over large item codebooks and enables fast retrieval within a unified architecture. We hope this additional analysis provides a comprehensive view of the reasoning overhead and helps clarify the practicality of our approach.

Q2: Short User History and Limited Context

Thank you for this insightful observation.

First, truncating user histories to the most recent actions is a common practice in the recommendation literature, primarily due to the structure of public datasets and for fair comparison across baselines.

Second, the LLM-based baselines all truncated sequences to 10 recent actions in their original experiment settings; in fact, we increased this window to 20 in our experiments to better retain semantic and sequential information, supporting stronger reasoning capabilities.

Nevertheless, we fully acknowledge such setting does not fully capture the complexity of real-world recommendation scenarios, where user histories are often much longer and richer. Addressing long-context modeling remains an important and challenging direction for the field.

Due to time and resource constraints, we were unable to reprocess all the datasets and conduct additional experiments with extended histories in this submission, but we do recognize this as an important aspect to explore in future work.

Q3: Adding Zero-shot Baseline

Thank you for this valuable suggestion. We agree that including a baseline using a pure pre-trained LLM—without any fine-tuning or RL—would help clarify the benefits of our approach.

We now report results for this "zero-shot" baseline. As shown below, raw pre-trained LLMs achieve very low performance on recommendation tasks. This further highlights the significant gains achieved by our proposed training scheme.

• Qwen

  	N@5	N@10	H@5	H@10
  CDs	0.000692	0.000692	0.00148	0.0022321
  Video Games	0.00016886	0.00023456	0.00028342	0.00048811
  Instruments	0.	0.00034	0.	0.000975

• Gemma

  	N@5	N@10	H@5	H@10
  CDs	0.00064089	0.00090592	0.00148	0.0022321
  Video Games	0.000317	0.00047956	0.00047237	0.00094473
  Instruments	0.	0.0001155	0.	0.0003253

Q4: No Discussion on Cold-start Scenario

Thank you for highlighting the importance of the cold-start scenario. In response, we conducted additional analysis on both “cold” (unseen user or item) and “warm” (seen user/item) settings. Results (NDCG@10) are shown below:

As shown, R²ec consistently outperforms SASRec in both cold and warm scenarios, especially under the more challenging cold-start setting. This demonstrates the strong generalization and robustness of our method.

Q5: Model Parameter Comparability

For all LLM-based baselines, the parameter count is kept strictly consistent by adapting LORA finetuning to ensure fair comparison. For traditional (non-LLM) baselines, model sizes are inherently smaller by design; for clarity, we will provide a summary table of model sizes in the appendix.

Q6: Sequence Embedding Extraction Strategies

Thank you for raising this important point and for sharing relevant recent works. Your concern strongly resonates with my own initial explorations—in the early stages of this project, I experimented with several sequence embedding strategies, but I did not observe substantial or consistent improvements over the last hidden state, so I adopted it as the default for simplicity and robustness.

Regard the papers. I found your recommended works extremely insightful—especially the discussion on enabling bidirectional encoding. Within the LLM + reasoning + embedding architecture, systematically investigating how to construct more effective embeddings is a highly promising direction for future research.

In direct response to your suggestion, I have now re-evaluated these embedding strategies in controlled experiments. The results are as follows:

These results suggested that the last hidden state currently yields the strongest performance, while mean pooling improves training stability (though, due to rebuttal constraints, I cannot show the full curves here). Special token strategies yield weaker performance in our setting, likely because we adopt LoRA fine-tuning and no large-scale optimization on the LLLM backbone, it is thus challenging for the model to effectively learn to generate and utilize newly introduced tokens for global sequence representation.

Summary

We thank the reviewer again for the detailed feedback, which has prompted us to significantly strengthen our empirical analysis and clarify methodological choices. In the revised manuscript, we will incorporate all of the above additions—including expanded latency analysis, the zero-shot baselines, cold-start results, reporting of model sizes, and a systematic comparison of embedding strategies. We believe these enhancements address your concerns and further demonstrate the robustness and transparency of our approach.

We thank the reviewer for the positive and constructive feedback. We appreciate your recognition of writing quality, code release, and experimental efforts.

There are also some questions including the terms, notations, and baselines. We believe these issues could be clarified in the rebuttal stage, which we address as follow.

1: On the use of the term "large recommender models"

Thank you for highlighting this concern. We would like to clarify our terminology as follows.

1) Definition and technical equivalence.

In prior literature [1,2], “large recommender models” generally refers to recommenders with large number of parameters, typically built on Transformer architectures. Although naming conventions differ, there is no fundamental technical gap between those models and ours; both rely on large-scale Transformer backbones, with the primary difference being whether the model is pretrained as a causal language model.

2) Motivation for terminology.

We use “large recommender models” to emphasize our focus on unified, end-to-end LLM recommenders that do not require extra modules, in contrast to dual frameworks like LangPTune and SLIM, where LLM is merely the reasoner, but not the recommender(see Related Work). In this sense, the more commonly used term “LLM-based recommender” cannot fully capture this architectural distinction, as many so-called “LLM-based” systems still rely on external modules.

Nevertheless, we appreciate your suggestion and will revise the manuscript to more clearly define this terminology in both the Introduction and Related Work sections, making explicit the architectural distinctions and ensuring our positioning is transparent to readers.

Q2: On notation clarity (reasoning token $o$, hidden state $h$ )

1. The “reasoning token” $o$ refers to the output token generated at each step during the model’s autoregressive reasoning process, as stated in line 192; and illustrated as purple blocks in Figure 1, as stated in the legend.
2. $h$ represents the model’s hidden states, not tokens, as stated in line 138-139, and illustrated in the legend of Figure 1.
3. To further improve clarity, we will revise the figure by adding a prominent arrow to clearly indicate the correspondence between the reasoning tokens and the illustration.

Q3: On the motivation and necessity of reasoning in recommendation

Thank you for raising this insightful question regarding the necessity of explicit reasoning in recommendation. We will derive this question into the following 3 parts, namely the research necessity, the empirical evidences, and the Reasoning Behavior Analysis.

1) Motivation and Research Necessity.

Generative recommendation (GR) built upon LLM backbones has already shown strong promise in both industry and academia. While recent advances demonstrate that reasoning is central to LLMs’ performance in complex tasks such as mathematics. Given that recommendation inherently involves intuitive reasoning—such as inferring user intent and abstracting preferences — it is both natural and timely to investigate whether reasoning can yield tangible gains in GR setting. We believe exploring this direction is crucial for advancing the capabilities and practical impact of GR systems.

2) Empirical Evidences.

1. Ablation studies (Section 5.3) demonstrated that our model consistently outperforms a variant trained solely with contrastive loss and without reasoning. This provides quantitative evidence that incorporating reasoning tokens leads to better recommendation performance.

2. In direct response to “is it really necessary to maintain the reasoning procedure”, we performed an additional ablation where R²ec is forced to output recommendations without generating any reasoning (denoted as w/o r), comparing with the reasoning one (denoted as w r). As shown below, removing reasonings consistently degrades performance across all domains.

   	N@5	N@10	H@5	H@10
   CDs w/o r	0.0222	0.0277	0.0327	0.0469
   CDs w r	0.0388	0.0457	0.0588	0.0804
   Video Games w/o r	0.0145	0.0184	0.0220	0.0341
   Video Games w r	0.0205	0.0271	0.0288	0.0532
   Instruments w/o r	0.0156	0.0178	0.0224	0.0309
   Instruments w r	0.0161	0.0203	0.0264	0.0397
   Electronics w/o r	0.0169	0.02160	0.0265	0.0414
   Electronics w r	0.0205	0.0252	0.0314	0.0462

These experiments directly confirm that explicit reasoning is not only useful, but necessary for maximizing the performance of generative recommendation models.

3) Reasoning Behavior Analysis.

To directly address the question of “how reasoning benefits the final recommendation, "we conducted an in-depth both qualitative and quantitative analysis of R²ec's reasoning patterns.

We first engaged in extensive discussions with colleagues to collaboratively analyze and summarize the diverse reasoning behaviors exhibited by R²ec. Through this process, we identified 7 typical reasoning strategies. Building on this framework, we then randomly sampled 100 reasoning outputs from each test set. These samples were annotated and analyzed through a combination of detailed human evaluation and automated assessment with GPT-4.1, ensuring both depth and consistency in our analysis.

1. Qualitative Analysis. We found that there are 7 different reasoning behaviors for recommendation, including:

   • Attribute Abstraction: Identifying underlying attributes of purchased items.
   • Pattern Clustering: Detecting frequent item patterns to form user-interest clusters.
   • Role-Based Reasoning: Inferring user roles or specific use-cases.
   • Temporal Reasoning: Using purchase timing and order for context.
   • Self-Explanation: Providing explicit rationales for recommendations.
   • Negative Preference Exclusion: Avoiding previously negatively-rated item types.
   • Multi-Objective Balancing: Managing trade-offs between conflicting user objectives.

2. Quantitative Analysis. To further substantiate our findings, we quantified the distribution of reasoning behaviors across different datasets. As noted, each reasoning output may exhibit multiple behaviors simultaneously. The table below reports the proportion of sampled reasoning outputs within each dataset that display each behavior:

   Dataset	Attribute Abstraction	Pattern Clustering	Role-Based Reasoning	Temporal Reasoning	Self-Explanation	Negative Preference Exclusion	Multi-Objective Balancing
   Instruments	0.95	0.99	0.98	0.22	1.00	0.07	0.72
   Games	1.00	0.98	0.92	0.08	0.91	0.02	0.51
   CDs	0.99	0.94	0.74	0.16	0.82	0.06	0.23
   Electronics	0.97	0.84	0.94	0.12	0.97	0.15	0.76

   Note: Each value reflects the proportion of sampled reasoning outputs within the corresponding dataset that demonstrate a given reasoning behavior.

   This quantitative results highlight several key trends. First, the distribution of reasoning behaviors differs across datasets, indicating that after training, the model self-adapts its reasoning patterns to fit the characteristics of each domain and its user-item interactions. Second, the variability in these proportions suggests that even within a single domain, R²ec flexibly selects different reasoning strategies based on the specific user sequence.

Overall, these results demonstrate that the model employs context-aware reasoning tailored to both the dataset and the individual user, contributing to improved recommendation performance and ultimately proving its necessity.

Q4: On SASRec performance and discrepancies with original papers

Thank you for pointing out this discrepancy. As described in Appendix C, our SASRec implementation uses cross-entropy (softmax) loss instead of binary cross-entropy (BCE), which significantly boost performance. We highlight that such findings are reasonable and consistent with previous findings [1], which reported that softmax-based SASRec can outperform many larger recommendation models. We do value your concerns and will highlight this implementation detail more clearly in the main text to avoid confusion.

Q5: On missing baselines

Thank you for highlighting the importance of recent works such as LLaRA[2] and SPRec[3].

To ensure broader coverage of recent advances, we supplemented our original baselines with three additional LLM‑based recommenders—SPRec, SDPO, and LLara. We note that LLaRA and SDPO are originally designed as rerankers in a candidate‑based setting, thus not comparable to our end‑to‑end generation framework. To provide fair comparison, we have adapted their implementations by removing the candidates prompts and instead performing constrained generation over the full item corpus, denoted as LLaRA* and SDPO*.

The updated comparison is shown below:

We hope these additions and clarifications adequately address your concern.

In summary, we sincerely thank the reviewer for the constructive feedback. We greatly appreciate your recognition of our writing, code release, and experiments, as well as your thoughtful suggestions on terminology, notation, and baselines. We have carefully addressed each concern,, providing detailed clarifications, new analyses, and additional experiments where possible. We hope these responses further strengthen the value and validity of our work.

[1] Are LLM-Based Recommenders Already the Best? Simple Scaled Cross-Entropy Unleashes the Potential of Traditional Sequential Recommenders [2] LLaRA: Large Language-Recommendation Assistant [3] SPRec: Self-Play to Debias LLM-Based Recommendation [4] On Softmax Direct Preference Optimization for Recommendation

Hi Reviewer ZdkM, if there are any points where our responses are unclear or have not fully addressed your concerns, we would be extremely grateful if you could let us know. This will allow us time to further clarify or revise our work accordingly. Your feedback is truly important to us, and we deeply appreciate your time, effort, and guidance.

Thank you again for your consideration.

Thank you for the thoughtful evaluation. We appreciate your recognition of the clarity, the significance of our research problem, and the strength of results.

We would like to address several misunderstandings on our contributions. Specifically, the novelty, the role of reasoning in recommendation, latency, and experiment result. Below, we provide a detailed response to your 5 primary questions.

Q1: Novelty.

While LLM reasoning advances make its adaptation to recommendation appear straightforward, our results show that simply applying standard RL and reasoning approaches fails in practice and critical design choices must be made. We summarize key contributions as (1) model architect and (2) training method.

1. Not a Classification Head — Architect considerations to couple reasoning and recommendation

   • Dual architect is suboptimal. Prior work typically adopts LLM as reasoner with recommender. This modular approach increases computation and decouples reasoning from recommendation, limiting effectiveness (see lines 37–47). Our goal is to design a unified architect that jointly accomplishes generative reasoning and discriminative recommendation.

   • Dense Embedding over Discrete Codebook. Generative recommendation (GR) often predict items as tokens from codebook, which restricts scalability, generalization to new items, and adds decoding cost. Since reasoning generation is already time-consuming, we use a dense item embedding paradigm for scalability and efficiency.

   • Item Semantics is crucial. While the simplest way to optimize dense embeddings in LLM is through a "classification head" (denoted as ClsH). Our experiments show that such design, with or without reasoning, performs poorly, under standard baselines like BigRec:

     	N@5	N@10	H@5	H@10
     ClsH w/o reasoning	0.0017	0.0019	0.0022	0.0037
     ClsH with reasoning	0.0025	0.0027	0.0030	0.0045
     BigRec	0.0025	0.0037	0.0045	0.0089

     Summary: After extensive early-stage exploration and careful considerations, our final architect (i) tightly couples reasoning and rec in a unified model, (ii) leverages scalable, efficient dense item representations, and (iii) enables the model to exploit and update item semantics throughout training.

2. RL for Rec is Nontrivial and Requires New Solutions

   • Reasoning Supervision is Practically Impossible. Annotated reasonings are feasible to collect at scale for math / code tasks. In recommendation, however, user-level annotated reasoning is impossible to obtain due to the diversity and subjectivity of user intents. RecPO is a fully self-supervised RL paradigm, learning reasoning purely from interaction logs—an essential, but highly challenging adaptation for this domain.

   • Reward Design is Nontrivial. NLP domains benefit from clear reward signals (e.g., corectness). In recommendation, ranking metrics are the only available rewards. We found that directly using such metrics leads to weak result, we thus proposed fused reward scheme to combine discrete ranking signals with continuous similarity (Sec. 4.2.2, Table 2), leading to much more effective and stable training.

   Summary: RecPO is the first to enable fully joint RL optimization of both reasoning and recommendation without reasoning annotation. By introducing encoding and utilization of item semantics to training objective, and optimizing with fused reward, our approach tightly couples reasoning and ranking, allowing reasoning to directly enhance recommendation.

Q2: Why reasoning process matters?

We agree that the reasoning of R²ec differs from mathematics domains. We thus conducted analyses regarding R²ec reasoning behaviors. By qualitatively annotating and clustering reasoning outputs, we found that R²ec typically leverages 7 reasoning strategies. To further substantiate the findings, we quantified the distribution of behaviors across different datasets as follow:

(Each value indicates the proportion of outputs exhibiting the behavior. Each output may contain multiple behaviors.)

This analysis demonstrates that R²ec improves recommendation quality by leveraging diverse and adaptive reasoning strategies, enabling it to capture richer user intent and item semantics beyond shallow co-occurrence patterns.

Q3 Real-time serving and latency.

We thank the reviewer for emphasizing real-time serving and latency concerns.

To address this, we conducted comprehensive latency evaluation, including comparison with both LLM-based Rec baselines and traditional recommenders, summarized as below.

These results demonstrate that R²ec achieves competitive efficiency among LLM-based recommenders, particularly when using modern inference frameworks. We attribute this efficiency in part to our use of scalable item embeddings, which avoids costly autoregressive token decoding over large item codebooks and enables fast retrieval within a unified architecture. We hope this additional analysis provides a comprehensive view of the reasoning overhead and helps clarify the practicality of our approach.

While R²ec is slower than lightweight models like SASRec, its unified reasoning–recommendation design brings substantial gains in recommendation quality, and aligns with the community’s ongoing shift toward more capable, explainable models. We believe such advances justify the increased computation and pave the way for efficient, practical LLM-based recommenders as the field continues to optimize for both performance and speed.

Q4: Baselines.

We included additional LLM-based baselines — SPRec, SDPO, and LLara. Since LLaRA and SDPO are originally designed as rerankers, , we adapted implementations by removing candidates prompts with constrained generation over item corpus, denoted as LLaRA* and SDPO*. The updated result is shown below:

As illustrates, R²ec maintains a substantial lead over the expanded set, further validating its effectiveness.

Q5: Concerns on experiment result.

Thank you for raising this important point and for your careful review. We take reproducibility and reliability very seriously, and address your concerns as follows:

1. Dataset Preprocessing. Our datasets are preprocessed following the D³ methodology (see Appendix for details), but without additional 5-core filtering. This results in a much higher proportion of cold-start cases—e.g., 39.9% of users in the CDs dataset and 34.7% in Instruments are cold-start. In such settings, generative baselines typically struggle, whereas R²ec benefits from richer semantics and demonstrates amplified improvements.

2. Baseline Implementation. For baselines such as LLaRA, BigRec, and D³, we strictly follow the original papers and use only the item titles as input features. This limited information further hinders their performance on cold-start items. In addtion, we reported relative improvements ((R²ec - baseline) / baseline), which can look inflated when baseline values are low.

3. Controlled Comparison Following your suggestion, we compared all methods using the same backbone (Gemma-2B) and LoRA adapter on the 2018 5-core CDs dataset. As shown below, R²ec consistently outperforms the version without reasoning (“w/o r”).

This result aligns with our expectations and prior findings: as user sequences become more informative, the relative advantage of reasoning diminishes, but the positive effect of reasoning remains robust and measurable.

4. Commitment to Reproducibility. We have provided our full codebase in supplementary materials to enable the community to verify and build on our work.

We hope these steps demonstrate the validity of our findings.

If any of our responses above remain unclear or do not fully resolve your misunderstandings, we would be truly grateful if you could let us know as soon as possible. This would allow us sufficient time to further discuss or improve our work. Your feedback is extremely valuable to us, and we deeply appreciate your guidance and support. Thank you so much for your time and patience.

Thank you for your time and for recognizing the value of the additional experiments. We will surely incorporate the new results into the final version.

We also respect that views on novelty can differ based on individual backgrounds and tastes. Our goal is simply to present our contributions: a) the unified reasoning-ranking architecture, b) label-free RL training (RecPO), and c) the stable fused reward, as clearly as possible and to learn from the feedback of the reviewers and ACs, minimizing misunderstandings as less as possible.

Besides novelty, may we confirm whether the other issues you raised, i.e., latency, baseline coverage, benchmarking result, and the benefit of reasoning—are now resolved in your view? Any final pointers would be greatly appreciated; with the rebuttal period now extended, we remain open to further explanations and discussions.

Thank you again for the constructive discussion.

We sincerely thank the reviewer for recognizing the intriguing nature of our study, the clear organization of the manuscript, and the promising experimental results. We also appreciate the constructive feedback regarding dataset diversity, the depth and nature of reasoning tokens, and the comparison to alternative ranking head architectures.

We address the above suggestions one by one as follows.

Q1: Dataset Diversity

Considering your suggestion on “more dissimilar verticals (e‑g., Fashion or Electronics)” and the limited rebuttal time, we have added the Electronics dataset and compared R2ec with the 2 advanced baselines, namely SASRec and. D³. The performance results are summarized as follows:

The results show that R²ecstill performs better on the new dataset, illustrating the cross-domain robustness on diverse domains.

Q2: Reasoning Behavior Analysis — How Does R²ec Reasoning to Recommend?

To better address your insightful question regarding the reasoning behavior, we conducted an in-depth both qualitative and quantitative analysis of R²ec's reasoning patterns.

We first engaged in extensive discussions with colleagues to collaboratively analyze and summarize the diverse reasoning behaviors exhibited by R²ec. Through this process, we identified 7 typical reasoning strategies. Building on this framework, we then randomly sampled 100 reasoning outputs from each test set. These samples were annotated and analyzed through a combination of detailed human evaluation and automated assessment with GPT-4.1, ensuring both depth and consistency in our analysis.

1. Qualitative Analysis. We found that there are 7 different reasoning behaviors for recommendation, including:

   • Attribute Abstraction: Identifying underlying attributes of purchased items.
   • Pattern Recognition & Clustering: Detecting frequent item patterns to form user-interest clusters.
   • Scenario/Role-Based Reasoning: Inferring user roles or specific use-cases.
   • Temporal & Sequential Reasoning: Using purchase timing and order for context.
   • Self-Explanation: Providing explicit rationales for recommendations.
   • Negative Preference Exclusion: Avoiding previously negatively-rated item types.
   • Multi-Objective Balancing: Managing trade-offs between conflicting user objectives.

   These distinct reasoning strategies help explain why and how reasoning aids recommendation. Large recommender model equipped with reasoning can better distill salient user preferences and intentions from noisy or complex behavioral data, abstract patterns beyond surface-level co-occurrence, and generate more contextually appropriate and user-aligned recommendations. This richer interpretive capacity enables the model to move beyond shallow heuristics, providing both higher accuracy and more interpretable results.

2. Quantitative Analysis. To further substantiate our findings, we quantified the distribution of reasoning behaviors across different datasets. As noted, each reasoning output may exhibit multiple behaviors simultaneously. The table below reports the proportion of sampled reasoning outputs within each dataset that display each behavior:

   Dataset	Attribute Abstraction	Pattern Recognition & Clustering	Scenario/Role-Based Reasoning	Temporal & Sequential Reasoning	Self-Explanation	Negative Preference Exclusion	Multi-Objective Balancing
   Instruments	0.95	0.99	0.98	0.22	1.00	0.07	0.72
   Games	1.00	0.98	0.92	0.08	0.91	0.02	0.51
   CDs	0.99	0.94	0.74	0.16	0.82	0.06	0.23
   Electronics	0.97	0.84	0.94	0.12	0.97	0.15	0.76

Note: Each value reflects the proportion of sampled reasoning outputs within the corresponding dataset that demonstrate a given reasoning behavior.

This quantitative results highlight several key trends. First, the distribution of reasoning behaviors differs across datasets, indicating that after training, the model self-adapts its reasoning patterns to fit the characteristics of each domain and its user-item interactions.

Second, the variability in these proportions suggests that even within a single domain, R²ec flexibly selects different reasoning strategies based on the specific user sequence.

Overall, these results show that the model does not follow a fixed template but instead employs context-aware reasoning tailored to both the dataset and the individual user, contributing to improved recommendation performance and interpretability.

Q3: Comparison with External Ranking Layer Baseline

We appreciate the suggestion to compare against an alternative approach—attaching an external ranking layer to the hidden state after the profile summary and training end-to-end via back-propagation.

To this end, we implemented and evaluated this design under two variants: (1) a ranking layer without explicit reasoning, and (2) a ranking layer with RecPO. As shown in the table below, both approaches result in substantially lower performance compared to our proposed tightly-coupled architecture:

These results underscore that a naively attached ranking layer—even with end-to-end training—fails to fully leverage the model’s reasoning capacity. In contrast, our unified design which utilizes item embedding enables deeper alignment between reasoning and recommendation, resulting in significantly better performance.

Q4: Larger-size LLMs

Thanks for your suggestion. We have conducted additional experiments with a larger backbone Gemma-9B-it, as you recommended. The results on our benchmark datasets are summarized as follows:

For comparison, we also present the results under Gemma-2B-it backbone from the original paper:

The results clearly demonstrate the consistent effectiveness of reasoning-augmented recommendation across model scales. With Gemma-9B-it, R²ec achieves a 21.7% improvement in N@5 compared to the version without reasoning, validating that the reasoning mechanism provides fundamental advantages regardless of parameter count. Notably, we observe that as model size increases from 2B to 9B parameters, the absolute performance gains from reasoning also increase, suggesting that larger models are more adept at leveraging reasoning capabilities to enhance recommendation quality.

Summary

In response to your suggestions, we have expanded our evaluation with additional datasets, conducted experiments with an external ranking layer, tested larger model architectures, and performed detailed analysis of reasoning behaviors in recommendation. These additions have strengthened rather than challenged our work, validating R²ec's robustness across diverse scenarios while more clearly demonstrating the value of reasoning-enhanced recommendation systems.

We sincerely hope our responses have addressed your concerns. If there are still any issues or if further clarification is needed, please do let us know at your earliest convenience, so we may have enough time to discuss and make improvements. Your feedback is very valuable to us, and we truly appreciate your guidance and support. Thank you so much for your time and patience.