Open challenges for agentic recommender systems: autonomy–control balance, external knowledge integration, and dynamic multimodal evaluation

Develop principled methods to (i) balance autonomy with controllability in agentic recommender systems, (ii) effectively incorporate external knowledge into recommendation pipelines, and (iii) design evaluation protocols for dynamic multimodal settings.

Background

The paper situates MemRerank within the emerging landscape of agentic recommender systems, where LLMs coordinate tools and memory. In summarizing broader perspectives, it highlights foundational challenges that remain unresolved.

Specifically, achieving the right autonomy–control trade-off, integrating external knowledge sources reliably, and devising robust evaluations for dynamic multimodal interactions are identified as open issues that the community has yet to solve.

References

Broader surveys on agentic recommender systems emphasise that balancing autonomy with controllability, incorporating external knowledge, and evaluating dynamic multimodal settings remain open challenges.

— MemRerank: Preference Memory for Personalized Product Reranking  (2603.29247 - Peng et al., 31 Mar 2026) in Introduction

Interestingly, yields stronger gains on fine-grained metrics (N@5 and H@5) than on H@20. We conjecture that continuous semantic assistance mainly enhances fine-grained recommendation, while adaptive control of assistant guidance could further improve broader candidate coverage. We leave this direction for future work.

— SCoRD: Semantic-Assisted Continual Retriever-Reranker Distillation for LLM-Based Recommendation  (2608.19998 - Baek et al., 20 Aug 2026) in Section 5, subsection “Main results,” subsection “Overall”