acceptodds
Under review as a conference paper at ICLR 2027

RRCM: Ranking-Driven Retrieval over Collaborative and Meta Memories for LLM Recommendation

Abstract

Large Language Models (LLMs) have emerged as a promising paradigm for next-generation recommender systems, offering strong semantic understanding and natural-language reasoning abilities. Despite recent progress, current LLM-based recommenders still face key challenges in constructing decision-relevant contexts from heterogeneous evidence. First, existing methods often rely on fixed context construction strategies: collaborative behavioral evidence and item-side metadata are typically incorporated through predefined prompts, static retrieval pipelines, or handcrafted injection mechanisms, making it difficult to determine what information is truly beneficial for each instance. Second, heterogeneous evidence introduces a severe context-efficiency bottleneck. Rich metadata and collaborative interaction records can quickly overwhelm the context window, while aggressive compression or heuristic filtering may discard fine-grained evidence critical for accurate recommendation. To address these challenges, we propose RRCM, a ranking-driven retrieval-and-reasoning framework over collaborative and metadata memories for LLM-based agentic recommendation. RRCM starts from a lightweight user-history context and learns whether to recommend directly, retrieve collaborative evidence, retrieve item metadata, or interleave both through reasoning. Both memories are represented in natural language and accessed through a unified retrieval interface, enabling flexible evidence acquisition without handcrafted CF injection or fixed retrieval rules. We optimize this memory-reading policy with an outcome-only ranking reward, instantiated using group relative policy optimization, so that retrieval decisions are directly driven by final top- recommendation quality. Extensive experiments show that RRCM significantly outperforms traditional baselines and diverse LLM-based recommendation approaches. Code is available at: https://anonymous.4open.science/r/RRCM-6233.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM-based Recommendation, Retrieval-Augmented Generation, Agentic Recommendation, Collaborative Filtering

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XeasOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

RRCM is a method for dynamically retrieving Collaborative Memory and Metadata Memory in LLM-based recommendation systems and training a retrieval policy optimized for final recommendation performance using GRPO. It demonstrated consistent performance improvements over existing methods across three datasets.

Critical strengths

Consistent Performance Improvements: It outperforms various recommendation baselines, and the effectiveness of each component was verified through ablation studies.

Effective Information Integration: It improves recommendation quality by selectively leveraging collaborative signals and item metadata.

Thorough Experimental Analysis: It provides a wide range of additional experiments covering retrieval behavior, long-tail performance, robustness, and computational cost.

Critical weaknesses

Limited Novelty: It is methodologically similar to existing RL-based adaptive retrieval methods such as Search-R1[1], and its originality beyond the extension to the recommendation domain is unclear.

Lack of Validation of Key Contributions: Without a strong static retrieval baseline using the same memory and computational budget, it is difficult to isolate the effectiveness of the adaptive policy itself.

Issues with Comparison Fairness: While RRCM accesses a large-scale external collaborative memory, some baselines cannot utilize the same information, which may lead to a conflation of the effects of additional information and methodological improvements.

[1] JIN, Bowen, et al. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516, 2025.

Questions for authors

How does RRCM fundamentally differ from existing RL-based adaptive retrieval frameworks such as Search-R1, beyond applying them to recommendation with heterogeneous memories?

Can the authors provide comparisons with stronger static or fixed-budget retrieval baselines using the same memory corpus, backbone, and training objective to isolate the contribution of the learned retrieval policy?

Given that RRCM accesses a substantially larger collaborative memory corpus than the training data available to some baselines, how can the authors distinguish improvements from additional information access versus the proposed methodology?

4 comments

Dear Reviewer 7ks9,

Thank you for your positive assessment of our information integration design, performance improvements, and thorough experimental analysis. We appreciate your constructive comments and the opportunity to clarify our experimental comparisons and methodological contribution.

W1/Q1. Methodological Novelty

Thank you for raising this point. We would like to clarify that our contribution is an application-driven methodology addressing an important challenge in LLM-based recommendation: selectively incorporating heterogeneous evidence under context-length and computational constraints. Practical recommender systems often contain thousands of item features and extensive interaction histories. Determining which evidence is useful for a particular recommendation, and how to fit it into a limited LLM context, remains an important unresolved problem. As the ICLR reviewer guidelines state, "Submissions bring value to the ICLR community when they convincingly demonstrate new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc)." We hope this clarifies our emphasis on methodological and empirical contributions to this application problem.

Our central insight is to make evidence acquisition a learnable part of the recommendation policy, guided by final ranking quality. Although similar components appear in retrieval-and-reasoning frameworks such as Search-R1, RRCM uses one interface for collaborative histories and item metadata and learns when and what to retrieve jointly with recommendation through end-to-end RL. In contrast, previous method K-RagRec uses a predefined strategy for metadata retrieval, while BinLLM and CoLLM incorporate collaborative information through learned embeddings. RRCM's significant performance gains over these methods demonstrate the value of adaptive evidence acquisition in this setting. The All-Evidence and Fixed-Retrieval controls in Appendix D.2 further show that these gains persist when access to both memories is held constant, with RRCM also using fewer inference tokens. We have revised the Introduction to clarify this contribution and its practical significance.

W2/Q2. Comparison with Retrieval Baselines

Thank you for this suggestion. We would first like to highlight that our evaluation already includes closely related approaches developed to retrieve and incorporate external evidence into LLM-based recommendation. K-RagRec retrieves item information from a knowledge graph, while BinLLM and CoLLM incorporate collaborative signals through learned representations. Moreover, the All-Evidence ablation in Appendix D.2 compares adaptive acquisition with fixed evidence provision using the same memories, retriever, backbone, reasoning format, and GRPO training.

To further isolate the contribution of the retrieval policy, we have additionally trained a Fixed-Retrieval model on Goodreads. We explicitly instruct the model to issue exactly one collaborative-memory query and one metadata-memory query before recommending, and adapt the format reward to penalize violations. The memories, retriever, backbone, reasoning format, ranking reward, and remaining GRPO settings are identical to those of RRCM. This comparison holds the evidence sources and training framework constant while replacing adaptive retrieval with a fixed query budget.

HR@5NG@5HR@10NG@10Inference Tokens
Fixed-Retrieval0.02060.01970.03510.0244862
RRCM0.02150.02110.03640.0264601

RRCM outperforms Fixed-Retrieval across all four metrics while using 30.3% fewer inference tokens. These results show that its gains persist against a model with the same evidence access and a higher inference cost. Together with All-Evidence, this comparison strengthens the evidence that learning selective retrieval improves both recommendation quality and efficiency. We have added these results and discussion to Appendix D.2.

Furthermore, we've added a control that trains a traditional sequential recommendation model, SASRec, on all available historical interactions, including those populating the collaborative memory. We compare it with RRCM trained with the same training set size in Table 13 (around 12,000). As shown below, RRCM still outperforms it across all four metrics. Thus, the performance advantage persists when SASRec is allowed to train on the interaction records available through RRCM’s memory. RRCM uses substantially fewer interactions for parameter updates while accessing the remaining behavioral evidence through retrieval. We have included these results in the revised Appendix D.1.

HR@5NG@5HR@10NG@10
SASRec0.08120.07060.13300.0862
RRCM0.09000.08400.16250.1098

W3/Q3. Baseline Comparison

We wish to highlight that we've already discussed this issue in detail in Appendix F.2. A central motivation of RRCM is that existing LLM recommenders struggle to incorporate extensive histories and rich metadata within limited context and computational budgets, often relying on signals encoded through training or fixed retrieval strategies. Simply appending the external memory to non-retrieval baselines would change their intended designs without resolving these constraints and can be impractical in many real-world settings. As such, adaptive access to external evidence is itself a methodological capability being evaluated. To distinguish its benefits from those of additional information alone, our evaluation already includes methods that retrieve and incorporate external evidence through their intended mechanisms, alongside controlled retrieval-policy comparisons.

Specifically, K-RagRec retrieves item information from a knowledge graph, while BinLLM and CoLLM incorporate collaborative signals through learned representations trained on exactly the same interaction records as RRCM’s collaborative memory. Thus, these comparisons assess different ways of representing and incorporating collaborative information with the same underlying data access. In contrast, RRCM learns when and what evidence to retrieve, allowing context construction to adapt to each recommendation instance.

Furthermore, the All-Evidence and Fixed-Retrieval experiments in Appendix D.2 directly control for access to both memory sources, using the same corpus, retriever, backbone, and ranking-based training objective. RRCM’s improvements over these controls support the effectiveness of its adaptive retrieval policy beyond the benefit of additional information. The ablations in Section 4.3 complement these comparisons by examining the contributions of collaborative memory, metadata memory, and reasoning. Together, these experiments distinguish the value of the evidence sources from the value of learning how to acquire and use them. We've further clarified the baseline data access and expanded this discussion in the revised paper, marked in red.

Dear Reviewer Xeas,

Thank you again for the time and care you have devoted to evaluating our paper and for your thoughtful comments. As the author–reviewer discussion period approaches its deadline, we would like to confirm that we have adequately addressed your remaining concerns.

In response to your feedback, we have clarified our methodological novelty and empirical positioning relative to general RL-based retrieval frameworks like Search-R1, added a controlled Fixed-Retrieval baseline using the exact same memory corpus and optimization setup to isolate the value of adaptive retrieval, and explicitly detailed how our collaborative-memory baselines and ablations ensure fair comparisons across data access.

Please let us know if any questions or concerns remain. We greatly value your feedback and would be glad to provide further clarification at any time.

Best regards,

The Authors

Dear Reviewer Xeas,

With one day remaining in the author–reviewer discussion period, we wanted to follow up briefly. We would appreciate any feedback you may have on our responses and revisions.

If any questions or concerns remain, we would be happy to address them anytime. If you feel that our revisions have addressed most of the concerns, we would be grateful if you would consider raising the score.

Thank you again for your time and thoughtful feedback!

Best regards,

The Authors

Reviewer_7ks9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper casts next-item recommendation with an LLM as adaptive evidence acquisition. Starting from a title-only user history, a Qwen3-1.7B policy decides whether to answer directly or to issue natural-language queries to a single dense retriever over two textual memories i.e., collaborative memory and meta memory. The policy is trained with GRPO on a weighted sum of top-n hit indicators plus a format penalty. Experiments on Goodreads, MovieLens and Amazon CDs and Vinyl report gains of roughly 9 to 21 percent (relative) over SFT, DPO, reasoning, RAG and CF-injection baselines, along with ablations, an All-Evidence control, cost analysis and behavior diagnostics.

Critical strengths
  1. Clear formulation. Context construction is learned rather than fixed: collaborative histories and item metadata are exposed as textual memories behind one retrieval interface, and the policy is trained only with a final ranking reward, with no handcrafted CF injection or retrieval rules. The resulting trajectories are interpretable.
  2. Transparent cost reporting. The paper reports training time, tokens, latency and index size, shows the quality–cost trade-off under a length penalty, and discusses realistic deployment.
  3. Broad analysis. The evaluation spans three domains and two backbones. Ablations reveal domain-dependent bottlenecks that match memory-source selection ratios (Table 7), and the paper includes robustness checks along with candid negative results.
Critical weaknesses
  1. Access to additional interaction data confounds the comparison with traditional baselines. Beyond the 4,096 shared training interactions, RRCM retrieves from 50K–133K held-out interaction records. Appendix F.2 frames this memory as the capability under evaluation, but the claim of outperforming traditional recommenders depends on it: with 3× training data, SASRec's Goodreads HR@10 rises from 0.0145 to 0.1094 (Table 13), about three times the best result in Table 1. The natural control is missing, namely a sequential recommender trained on the interactions that populate the collaborative memory. The training data for BinLLM/CoLLM's collaborative embeddings is also unspecified.
  2. Novelty is overstated relative to RL-trained agentic recommenders. The training recipe (think/tool-call/answer tags, Flat E5 retrieval, GRPO with retrieved-token masking) closely follows Search-R1 [1]. AgenticRec [2] already trains an LLM recommender with GRPO and an NDCG-based outcome reward to interleave reasoning with tools that return item information and collaborative evidence. Thus, Appendix C's characterization of prior agentic recommenders as training-free is inaccurate, and the contribution should be positioned against this line of work.
  3. The evidence for learned selective retrieval is indirect. The paper reports a mean of 0.93 retrievals per instance, but not the distribution: how often the policy skips retrieval, and whether the skipped instances are those where retrieval would not help. The declining curves in Figure 2 are equally consistent with the generic response shortening seen under GRPO. The All-Evidence control changes both the selection mechanism and the evidence content, so it does not isolate adaptivity. An identically trained fixed-budget control (e.g., always one collaborative and one metadata query) would.

Reference

[1] Search-r1: Training llms to reason and leverage search engines with reinforcement learning

[2] AgenticRec: A Recommendation-Oriented Agentic Framework with Progressive Tool-Integrated Reasoning Optimization

Questions for authors

Please refer to the weaknesses.

5 comments

Dear Reviewer 7ks9,

Thank you for your positive assessment of our motivation, formulation, cost reporting, and broad analysis. We appreciate your constructive comments and the opportunity to clarify our experimental comparisons and methodological contribution.

W1. Traditional Baseline Comparison

Thank you for raising this valuable point. First, we'd like to highlight that, as described in the Introduction, this paper primarily addresses challenges in LLM-based recommendation, particularly adaptive context construction and the context-efficiency bottleneck. We therefore focus our main comparisons on LLM-based recommenders and follow the training-data protocol adopted in prior studies such as SPRec and BIGRec. This provides a consistent setting for comparing LLM adaptation methods.

Nevertheless, we agree that a more detailed comparison with traditional recommenders would be helpful. Following your suggestion, we have added a control that trains SASRec on all available historical interactions, including those populating the collaborative memory. We compare it with RRCM trained with the same training set size in Table 13 (around 12,000). As shown below, RRCM still outperforms it across all four metrics. Thus, the performance advantage persists when SASRec is allowed to train on the interaction records available through RRCM’s memory. RRCM uses substantially fewer interactions for parameter updates while accessing the remaining behavioral evidence through retrieval. We have included these results in the revised Appendix D.1.

HR@5NG@5HR@10NG@10
SASRec0.08120.07060.13300.0862
RRCM0.09000.08400.16250.1098

Regarding BinLLM and CoLLM, we follow their original training procedures and train their collaborative embeddings on exactly the same interaction records that constitute RRCM’s collaborative memory. BinLLM converts these embeddings into binary textual sequences for inclusion in the prompt, while CoLLM maps them into the LLM’s input embedding space. Both baselines therefore access the same underlying collaborative data through their intended mechanisms. We have made this setting explicit in the revised experimental setup.

W2. Novelty Statement

Thank you for suggesting these relevant works. First, we'd like to highlight that our contribution is an application-driven methodology addressing an important challenge in LLM-based recommendation: selectively incorporating rich item features and extensive interaction histories under context-length and computational constraints. Our central insight is to make evidence acquisition a learnable part of the recommendation policy, guided by final ranking quality. As a result, we deem an agentic RL framework with retrieval and GRPO as a natural and well-motivated choice. Although similar components also appear in retrieval-and-reasoning frameworks such as Search-R1, RRCM’s contribution lies in formulating and validating their use for unified, selective acquisition of heterogeneous recommendation evidence. We've revised the Introduction to clarify this contribution and its practical significance.

We thank the reviewer for highlighting AgenticRec, a relevant recent work we inadvertently missed during our literature review. Its preprint appeared earlier this year, followed by publication in Findings of EMNLP in October, after the ICLR submission deadline. We have added a comparison in Appendix C and revised our characterization of prior agentic recommenders. Both methods use GRPO to interleave reasoning and evidence acquisition, but differ in three aspects. First, AgenticRec uses four specialized tool categories, including pregenerated profiles and retrieval based on trained SASRec embeddings. RRCM retrieves textual histories and metadata through one unified interface, avoiding these additional preprocessing and training pipelines. Second, AgenticRec combines trajectory optimization and a tool-call bonus with subsequent hard-pair preference refinement. RRCM jointly optimizes retrieval and recommendation through a single end-to-end RL stage guided by final ranking quality, without tool-use incentives or pairwise refinement. Finally, AgenticRec ranks 20 supplied candidates, whereas RRCM generates recommendations grounded against the full catalog.

W3. Selective Retrieval Analysis

Thank you for these valuable suggestions. We'd first like to highlight that Appendix D (in the revised version, previously C.1) already provides several complementary analyses supporting the effectiveness of selective retrieval. We examine memory-source preferences, assess the contribution of retrieved evidence, evaluate query quality, retrieval relevance, and sensitivity to retrieval noise. Together, these analyses show how the policy selects between evidence sources and how the retrieved information contributes to recommendation quality. Following your suggestions, we've added more analyses of retrieval skipping and a fixed-retrieval control.

First, we report the proportions of test instances with zero, one, or multiple retrieval calls:

DatasetNo retrievalOne retrievalMultiple retrievals
Goodreads23.4%56.1%20.4%
MovieLens18.8%52.3%28.9%

We then conduct an evidence-injection diagnostic on instances where RRCM chooses to skip retrieval. For each such instance, we explicitly instruct the model to issue one collaborative-memory query and one metadata-memory query, and incorporate the retrieved information into its context before making the final recommendation. We compare the resulting recommendation with the original trajectory and report the percentage of instances where this forced-retrieval intervention improves the recommendation metrics.

(Ratio with Improve)NG@5HR@5NG@10HR@10
Goodreads3.4%2.6%4.3%3.0%
MovieLens2.1%1.6%2.7%2.4%

As shown above, additional retrieval improves recommendation performance in only a very small proportion of skipped instances. This supports the policy’s ability to identify cases where retrieving further evidence offers limited or no benefit. We've included these results in Appendix D.

Second, following your suggestion, we train a Fixed-Retrieval model by explicitly instructing it to issue exactly one collaborative-memory query and one metadata-memory query before recommending. We incorporate this instruction into the prompt and adapt the format reward to penalize violations. The memories, retriever, LLM backbone, reasoning format, ranking reward, and remaining GRPO settings are identical to those of RRCM. This provides a direct comparison between a learned selective policy and a fixed retrieval strategy trained with the same evidence sources and optimization framework.

HR@5NG@5HR@10NG@10Inference Tokens
Fixed-Retrieval0.02060.01970.03510.0244862
RRCM0.02150.02110.03640.0264601

On Goodreads, RRCM outperforms Fixed-Retrieval across all four metrics while using 30.3% fewer inference tokens. These results strengthen the evidence that learned selective retrieval improves both recommendation quality and efficiency, beyond the declining retrieval counts observed during training. We've added this comparison to Appendix D.2.

Dear Reviewer 7ks9,

Thank you again for the time and care you have devoted to evaluating our paper and for your thoughtful comments. As the author–reviewer discussion period approaches its deadline, we would like to confirm that we have adequately addressed your remaining concerns.

In response to your feedback, we have added a controlled comparison with SASRec trained on all available historical interactions and clarified the training data parity for BinLLM and CoLLM, positioned our method against recent agentic recommenders such as AgenticRec, and provided additional selective-retrieval evidence including retrieval-count distributions, forced-retrieval diagnostics on skipped instances, and a controlled Fixed-Retrieval baseline.

Please let us know if any questions or concerns remain. We greatly value your feedback and would be glad to provide further clarification at any time.

Best regards,

The Authors

Reviewer_SiMHOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a ranking-driven retrieval and reasoning framework, RRCM, for LLM-based agentic recommendation. RRCM integrates two external memory sources, and dynamically decides whether additional information is required, formulates retrieval queries, and interleaves retrieved evidence with reasoning before generating a recommendation. Group Relative Policy Optimization is used to optimize the model with a reward based on top-k recommendation performance. The main contribution is the integration of adaptive retrieval over heterogeneous recommendation memories with outcome-driven reinforcement learning. Extensive experiments on three public recommendation datasets show improvements over traditional sequential recommendation methods and several LLM-based baselines.

Critical strengths
  1. This paper presents a simple yet effective solution for incorporating heterogeneous evidence into LLM-based recommendation while managing context length and retrieval costs.
  2. The idea of integrating collaborative and metadata memories through a unified retrieval interface and optimizing their use with a recommendation-specific ranking reward is interesting and has been demonstrated to be effective for sequential recommendation.
  3. The authors conducted comprehensive experiments to evaluate the proposed method from multiple perspectives, including recommendation performance, computational cost and latency, sensitivity to noise, and scalability.
Critical weaknesses
  1. The technical novelty is limited, with most components being straightforward combinations of existing techniques. While this integration has practical value, the authors should more clearly point out what methodological insight extends beyond simply combining existing components, and how it differs methodologically from recent closely related work such as RANK-GRPO.
  2. The main experimental comparisons are conducted with relatively small LLM backbones. While Appendix B.1 includes an additional 3B model, the conclusions about generalizability to larger models and practical recommendation environments remain limited.
  3. The experimental evaluation lacks comparisons with closely related adaptive retrieval methods, such as Self-RAG and Adaptive-RAG.
Questions for authors
  1. Have the authors observed any differences in retrieval preferences between users with short interaction histories and those with longer histories? For example, does the model rely more heavily on collaborative memory when user histories are sparse?
  2. Could the authors provide further analysis of how retrieval behavior varies with item popularity? In particular, does RRCM retrieve metadata more frequently for long-tail items than for popular items?
4 comments

Dear Reviewer SiMH,

Thank you for your positive assessment of the presentation, motivation, unified memory design, and experimental comprehensiveness of our work. We appreciate your thoughtful and constructive comments, which help us further clarify and improve the paper. We hope the responses below can help to address your concerns.

W1. Technical Novelty

Thank you for raising this point. We'd like to highlight that our contribution is an application-driven methodology addressing an important challenge in LLM-based recommendation: selectively incorporating heterogeneous evidence under context-length and computational constraints. Practical recommender systems often contain thousands of item features and extensive interaction histories, making their effective integration into a limited LLM context an important unresolved problem. As the ICLR reviewer guidelines state, “Submissions bring value to the ICLR community when they convincingly demonstrate new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc).” We hope this clarifies our emphasis on methodological and empirical contributions to this application problem.

Our central insight is to make evidence acquisition a learnable part of the recommendation policy, guided by final ranking quality. An agentic RL framework with retrieval and GRPO is a natural fit because it allows the model to learn when to seek additional evidence, which memory to query, and how to use the results when recommending. RRCM achieves significantly better performance than K-RagRec’s predefined metadata-retrieval strategy and the collaborative embedding approaches of BinLLM and CoLLM, demonstrating the practical value of learning these decisions jointly. The All-Evidence and Fixed-Retrieval controls in Appendix D.2 further test this insight: with access to the same memories, RRCM achieves better recommendation quality while using fewer inference tokens. We have revised the Introduction to highlight this insight and its practical significance, with changes marked in red.

Regarding Rank-GRPO, it is designed to resolve a different problem from ours: improving credit assignment and policy updates for list-structured outputs. RRCM addresses the construction and dynamic selection of heterogeneous external context. Since Rank-GRPO does not itself provide a mechanism for adaptive memory acquisition, it is not a direct baseline for the problem studied here. We have expanded Section 2 to clarify this distinction.

W2. Small LLM Backbones

Thank you for this suggestion. We originally selected our backbone sizes with reference to previous research and industrial applications, while considering the computational budget of our experiments and the cost requirements of potential deployment. For example, SPRec uses Llama-3.2-1B-Instruct as backbone, while LatentR3, which similarly employs SFT initialization followed by GRPO-based optimization, evaluates Qwen2.5-1.5B and Qwen2.5-3B. Furthermore, the influential OneRec Technical Report [1] from Kuaishou evaluates generative recommendation models with 0.015B, 0.121B, 0.935B, and 2.633B parameters, illustrating the relevance of this model-size range to industrial recommendation.

Recommender systems serving large user populations often require frequent updates to accommodate new items and evolving preferences. Given the resulting training and serving costs, together with the additional expense of RL training and limited budget for academic researchers like us, we chose Qwen3-1.7B and Llama-3.2-3B to balance experimental feasibility and practical relevance.

Nevertheless, following the reviewer's suggestion, we have conducted additional experiments with Qwen2.5-7B on Goodreads. As shown below, RRCM continues to outperform the compared methods, demonstrating that its benefits extend beyond the original 1.7B and 3B backbones. We have included these results in the updated Appendix B.2.

MethodHR@5NDCG@5HR@10NDCG@10
BIGRec0.01880.01780.02830.0211
LatentR30.02310.02110.03200.0238
K-RagRec0.02370.02150.03280.0245
BinLLM0.02250.02210.03170.0255
RRCM0.02550.02320.03720.0273

W3. Comparisons with Other Methods

Thank you for suggesting these methods. We would first like to highlight that our evaluation already includes closely related approaches specifically developed to retrieve and incorporate external evidence into LLM-based recommendation. K-RagRec retrieves item information from a knowledge graph, while BinLLM and CoLLM incorporate collaborative signals through learned representations. These methods directly address the recommendation setting and evidence sources central to our work. Our All-Evidence ablation further provides a controlled comparison of adaptive acquisition against fixed evidence provision using the same memories, backbone, and GRPO training.

Self-RAG and Adaptive-RAG are relevant to the broader adaptive-retrieval literature, but their original formulations address knowledge-intensive generation and question answering. A meaningful comparison requires adapting their task formulation, supervision, and evidence access to personalized recommendation. Our primary comparisons therefore focus on established recommendation methods that directly address our setting. This distinction also motivates RRCM’s use of final ranking performance to guide the acquisition of collaborative and metadata evidence using RL.

Nevertheless, to address the reviewer's suggestion, we have additionally evaluated a recommendation-adapted version of Adaptive-RAG. We retain its classifier-based selection among no-retrieval, single-step, and iterative-retrieval strategies, adapt and SFT the output for recommendation, and provide retrieval-enabled routes with access to the same two memories. Using the same backbone and evaluation protocol on Goodreads, RRCM achieves better recommendation performance, supporting the effectiveness of jointly learning retrieval and reasoning through the final ranking reward. We have added the adaptation details and results to Appendix D.2 and expanded the related-work discussion.

HR@5NDCG@5HR@10NDCG@10
Adaptive-RAG (adapted)0.01830.01590.02790.0193
RRCM0.02150.02110.03640.0264

Q1. Retrieval across User-History Lengths

Thank you for this valuable suggestion. We have conducted an additional analysis comparing users with the shortest 30% and longest 30% of interaction histories among the test sets within each dataset, measured using interactions available before prediction. We report the collaborative-memory retrieval rate, defined as the percentage of test instances containing at least one collaborative-memory query.

Short history rateLong history rate
Goodreads35.3%24.5%
MovieLens42.7%26.9%
CDs & Vinyl50.7%33.6%

As shown above, RRCM accesses collaborative memory more frequently when user histories are sparse. This is intuitively reasonable: longer histories provide richer preference information directly for LLM, whereas shorter histories may need more collaborative evidence from other users to supplement the available context. This pattern supports the motivation for adaptive evidence acquisition. We have added this analysis to the revised Appendix D.

Q2. Retrieval across Item-Popularity Levels

Thank you for raising this point. Appendices C.4 and F.1 already examine recommendation performance across item-popularity groups. Following your suggestion, we have extended this analysis to metadata-retrieval behavior. Specifically, we group items appearing in the LLM’s test-time input contexts into the bottom 30% and top 30% of popularity, measured by training interaction frequency. For each group, we calculate the proportion of item occurrences for which the model issues at least one metadata-memory query in the corresponding test instance.

As shown above, RRCM queries metadata more frequently per occurrence of long-tail items than popular items. This suggests that the model seeks additional semantic information for niche items more frequently, which can be less familiar or sparsely represented in behavioral data. We have added this analysis to the revised Appendix D.

DatasetLong-tail item ratePopular item rate
Goodreads15.2%7.8%
MovieLens12.6%6.7%
CDs & Vinyl6.8%3.9%

References

[1] Zhou, Guorui, et al. "Onerec technical report." arXiv preprint arXiv:2506.13695 (2025).

Dear Reviewer SiMH,

Thank you again for the time and care you have devoted to evaluating our paper and for your thoughtful comments. As the author–reviewer discussion period approaches its deadline, we would like to confirm that we have adequately addressed your remaining concerns.

In response to your feedback, we have clarified our application-driven methodological contribution and its distinction from Rank-GRPO, conducted additional experiments using a larger Qwen2.5-7B backbone as well as a recommendation-adapted Adaptive-RAG baseline, and added detailed analyses of retrieval preferences across user-history lengths and item-popularity levels.

Please let us know if any questions or concerns remain. We greatly value your feedback and would be glad to provide further clarification at any time.

Best regards,

The Authors

Dear Reviewer SiMH,

With one day remaining in the author–reviewer discussion period, we wanted to follow up briefly. We would appreciate any feedback you may have on our responses and revisions.

If any questions or concerns remain, we would be happy to address them anytime. If you feel that our revisions have addressed most of the concerns, we would be grateful if you would consider raising the score.

Thank you again for your time and thoughtful feedback!

Best regards,

The Authors

Comments on the paper

1 comment

Dear ACs and Reviewers,

Thank you for your careful evaluation and constructive feedback. We appreciate the reviewers’ recognition of RRCM’s effective integration of heterogeneous evidence, consistent performance improvements, and extensive analyses of retrieval behavior and computational cost. Following your suggestions, we have clarified the methodological contribution, strengthened the controlled comparisons, and expanded the experimental evaluation. We summarize the main responses and revisions below.

1. Methodological Contribution

We would like to highlight that our contribution is an application-driven methodology addressing an important challenge in LLM-based recommendation: selectively incorporating extensive interaction histories and rich item metadata under context-length and computational constraints. Our central insight is to make evidence acquisition a learnable part of the recommendation policy, guided by final ranking quality. An agentic RL framework with retrieval and GRPO is therefore a natural fit. Although similar components appear in frameworks such as Search-R1, RRCM learns when and what to retrieve from both memories jointly with the final recommendation. RRCM significantly outperforms K-RagRec, which retrieves metadata using a predefined strategy, and BinLLM and CoLLM, which incorporate collaborative signals through separately trained embeddings. These comparisons demonstrate the value of learning how to acquire and use evidence through end-to-end RL.

We have strengthened the Introduction and related-work discussion to clarify this contribution. We distinguish Rank-GRPO’s focus on rank-level credit assignment from our study of adaptive memory acquisition. We have also revised our characterization of prior agentic recommenders and added a comparison with AgenticRec that clarifies three aspects of RRCM: its unified retrieval interface, joint optimization in a single RL stage, and full-catalog recommendation setting.

2. Selective Retrieval Evidence

Our original All-Evidence comparison already controls the memories, retriever, backbone, reasoning format, and GRPO training while comparing adaptive acquisition with fixed evidence provision. Following the reviewers’ suggestions, we additionally trained Fixed-Retrieval, explicitly instructing the model to issue one collaborative-memory query and one metadata-memory query before recommending, with the same ranking reward and remaining GRPO settings.

On Goodreads, RRCM outperforms Fixed-Retrieval across all four metrics while using 30.3% fewer inference tokens. Together with All-Evidence, this comparison supports the value of learning selective retrieval beyond simply accessing additional information. We have included these results in Appendix D.2.

We also report the retrieval-count distribution and examine cases where RRCM skips retrieval. Forcing additional retrieval improves only a small proportion of these instances. Further analyses show that the model accesses collaborative memory more frequently for shorter user histories and queries metadata more frequently for long-tail items. These findings provide additional evidence of context-dependent retrieval behavior.

3. Comparison Fairness

We have clarified that BinLLM and CoLLM train their collaborative embeddings on the same interaction records as RRCM’s collaborative memory. Together with K-RagRec and the controlled retrieval comparisons above, these baselines help distinguish the benefits of evidence access from how that evidence is incorporated.

We additionally trained SASRec on all available historical interactions, including the collaborative-memory records. RRCM retains higher performance across all four metrics, using much fewer interactions for parameter updates while accessing external memory. These results appear in Appendix D.1, with data-access clarifications added to the experimental setup and fairness discussion.

4. Expanded Evaluation and Comparison

Our original backbone choices reflected prior recommendation research, industrial applications, and practical training and serving costs. We have now added Qwen2.5-7B experiments, where RRCM retains its advantage over the compared methods across all four metrics, extending the evidence beyond the original 1.7B and 3B backbones. We also evaluated a recommendation-adapted Adaptive-RAG baseline with access to the same two memories. RRCM achieves higher performance across all four metrics. These additions are reported in Appendices B.2 and D.2, respectively.

We have incorporated these clarifications and additional experiments into the revised manuscript, with textual revisions marked in red. We hope the strengthened comparisons and analyses address the main concerns and support a positive reassessment of our work.

We would be glad to provide further clarification at any time. Thank you again for your time and constructive suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.