BEGINNING OF COMMENTS TO THE AUTHOR(S)
****************************************
Recommended Decision by Associate Editor: Recommendation #1: Accept with minor revision (Revision time limit: 30 days)
Comments to Author(s) by Associate Editor:
Associate Editor / Guest Editor
Comments to the Author (Please be specific):
This manuscript addresses a timely and important topic—agentic recommender systems—and provides a well-structured and generally clear survey of LLM-based agent components.
That said, the reviewers have identified several areas where clarification and refinement would strengthen the contribution.
* First, the authors should more clearly position this survey relative to existing work.
* Second, the conceptual distinction between agentic recommender systems and conventional recommender systems should be sharpened. * Third, while the survey provides a comprehensive overview of LLM agent architectures, it currently resembles a general LLM agent survey. The authors are encouraged to place greater emphasis on recommendation-specific challenges and adaptations.
* Finally, the paper would benefit from a more in-depth discussion of scenario suitability—i.e., when agentic recommender systems are most appropriate.
Overall, the paper is well-written and addresses an important emerging topic. With the above revisions, it will make a valuable contribution to the literature.
********************
Comments to Author(s) by Reviewer(s):
Reviewer: 1
Recommendation: Accept with minor revision (Revision time limit: 30 days)
Comments:
This survey addresses the timely topic of agentic recommender systems and provides a well-structured overview of LLM-based agent components, including profiling, planning, memory, and action. The paper is generally clear and demonstrates solid coverage of related work across recommender systems and LLM agents.
However, the survey would benefit from clearer differentiation from existing work. Several surveys on LLM-powered agents and LLMs for recommendation have already been published, and the current manuscript overlaps substantially with these efforts. The claimed novelty as the first survey on agentic recommender systems is therefore not yet sufficiently substantiated.
Additionally, some of the core abilities attributed to agentic recommender systems—such as anticipation, initiative, and self-improvement—are already present in conventional recommender systems, while strategic planning closely resembles existing multi-stage architectures. As a result, the conceptual distinction between agentic and conventional recommender systems remains unclear.
Finally, the paper largely follows the structure of general LLM agent surveys and places limited emphasis on recommendation-specific challenges and adaptations, such as efficiency constraints, personalization-oriented memory design, and scenario suitability. A deeper discussion of when and where agentic recommender systems are most beneficial would significantly strengthen the survey’s contribution.
Overall, while the topic is important and the coverage is broad, clearer positioning and stronger recommendation-specific analysis are needed to enhance the impact of this survey.
Additional Questions:
How relevant is this manuscript to the readers of this journal? Please explain your rating in the Detailed Comments section.: Relevant
Are the title, running head, abstract, and keywords appropriate? Please elaborate in the Detailed Comments section.: Yes
Does the manuscript contain sufficient and appropriate references? Please elaborate in the Detailed Comments section.: References are sufficient and appropriate
Does the introduction state the objectives of the manuscript in terms that encourage the reader to read on? Please explain your answer in the Detailed Comments section.: Could be improved
Please rate and comment on the readability of this manuscript in the Detailed Comments section.: Easy to read
How would you rate the organization of the manuscript? Please elaborate in the Detailed Comments section.: Satisfactory
Is the manuscript focused? Please elaborate in the Detailed Comments section.: Could be improved
Is the length of the manuscript appropriate for the topic? Please elaborate in the Detailed Comments section.: Satisfactory
Please summarize what you view as the key point(s) of the manuscript and the importance of the content to the readers of this periodical.: This paper surveys recent advances in the field of agentic recommender systems. Starting from the design paradigm of LLM-based agents, the authors systematically introduce the design methodologies of key agent components, including profiling, planning, memory, and action. The coverage of related work is relatively comprehensive. However, the core content does not differ substantially from existing surveys on general-purpose LLM agents. The paper would benefit from placing greater emphasis on the distinctive characteristics of recommendation scenarios and from discussing the necessity and suitability of agent-based paradigms across different types of recommender systems.
What do you see as the three strongest points of this manuscript?: 1. Agentic recommender systems constitute a highly timely and forward-looking research topic, and a survey in this direction is indeed much needed by the community.
2. The authors provide an extensive and well-informed review of related work, demonstrating strong familiarity with both recommender systems and LLM-based agents.
3. The paper is generally well written, with a clear structure and organization that facilitates readability.
What do you see as the three weakest points of this manuscript?: 1. The paper claims to be the first survey on agentic recommender systems; however, related surveys have already been published. In addition, there has been a growing number of surveys on LLMs for recommendation in recent years. The authors should more clearly articulate the differences between this work and existing surveys, such as "A Survey on LLM-powered Agents for Recommender Systems" and "When Large Language Models Meet Personalization: Perspectives of Challenges and Opportunities".
2. Among the four core abilities discussed in Section 5.6, three—Anticipation, Initiative, and Self-Improvement—are already present in conventional recommender systems to a considerable extent. Regarding Strategic Planning, if modern multi-stage recommender systems are interpreted as a form of planning process, there is also a conceptual correspondence. From this perspective, it is not entirely clear how agentic recommender systems can be clearly distinguished from conventional recommender systems.
3. At present, the survey closely resembles general LLM agent surveys, with a primary focus on the modular components of LLM agents and the methods used within each module, many of which overlap significantly with prior agent surveys. Given that the focus is on recommender agents, the paper should place more emphasis on the adaptations and innovations required when deploying LLM agents in recommendation settings, such as efficiency-oriented designs to meet latency constraints or memory architectures tailored to enhance personalization.
4. The survey would benefit from a deeper discussion of which types of recommendation scenarios are most suitable for agentic recommender systems. Agents are fundamentally designed to improve information acquisition efficiency. However, in short-video platforms oriented toward entertainment, users can already access content streams with very low interaction costs, which may limit the practical value of agents. In contrast, in domains such as news recommendation—characterized by lengthy, dense, and complex information—users may be more inclined to adopt agent-based interfaces to accelerate information consumption and decision-making.
Please help ACM create a more efficient time-to-publication process: Using your best judgment, what amount of copy editing do you think this paper needs?: Moderate
Most ACM journal papers are researcher-oriented. Is this paper of potential interest to developers and engineers?: No
Reviewer: 2
Recommendation: Accept with minor revision (Revision time limit: 30 days)
Comments:
Issues to be addressed:
1. The term “multimodal” in the title is not often seen in the text. I suggest either removing it or incorporating more discussion about it.
2. The last paragraph in the introduction overlaps with the one before bullet points, both offering a summary of what follows up.
3. A discussion about the relevance and difference between this survey and [1] should be incorporated into section 3, since the latter is also about LLM-based agents for RS.
4. There is a comparison between different agents for RS in Table 3. It would be good to show the datasets that they utilized, since the environment of traditional RS and agentic RS is different, with the latter emphasizing interactivity.
[1] Zhu, Xi, et al. "Recommender systems meet large language model agents: A survey." Foundations and Trends® in Privacy and Security 7.4 (2025): 247-396.
Additional Questions:
How relevant is this manuscript to the readers of this journal? Please explain your rating in the Detailed Comments section.: Very Relevant
Are the title, running head, abstract, and keywords appropriate? Please elaborate in the Detailed Comments section.: Yes
Does the manuscript contain sufficient and appropriate references? Please elaborate in the Detailed Comments section.: References are sufficient and appropriate
Does the introduction state the objectives of the manuscript in terms that encourage the reader to read on? Please explain your answer in the Detailed Comments section.: Yes
Please rate and comment on the readability of this manuscript in the Detailed Comments section.: Easy to read
How would you rate the organization of the manuscript? Please elaborate in the Detailed Comments section.: Satisfactory
Is the manuscript focused? Please elaborate in the Detailed Comments section.: Satisfactory
Is the length of the manuscript appropriate for the topic? Please elaborate in the Detailed Comments section.: Satisfactory
Please summarize what you view as the key point(s) of the manuscript and the importance of the content to the readers of this periodical.: This manuscript is a survey about agentic recommender systems. Although there are many surveys about LLM-based RS, it differs from them in the perspective of LLM-based agents. The motivation is quite clear: traditional RS models are usually domain-specific, overfit to users’ interaction history, and lack interactivity, scrutability and transparency. LLM-based agents have the potential to resolve these issues owing to the following capabilities: anticipation, initiative, planning, and self-improvement. The authors mathematically defined what agent-based RS look like, surveyed recent papers published on top-tier conferences and journals, and outlined directions that are worth further investigation. I belive this manuscript can offer valuable insights for the RS community.
What do you see as the three strongest points of this manuscript?: The manuscript is well written and easy-to-follow.
A formal definition of agentic recommender systems is given.
It provides valuable insights for the community regarding the future development of agentic RS.
What do you see as the three weakest points of this manuscript?: As an import aspect of the survey, “multi-modality” seems to be less stressed.
A highly relevant and similar survey did not appear in the discussion.
A section about datasets for agentic recommender systems is missing.
Please help ACM create a more efficient time-to-publication process: Using your best judgment, what amount of copy editing do you think this paper needs?: None
Most ACM journal papers are researcher-oriented. Is this paper of potential interest to developers and engineers?: Maybe
****************************************
END OF COMMENTS TO THE AUTHOR(S)