Deep Research for Recommender Systems

Abstract: The technical foundations of recommender systems have progressed from collaborative filtering to complex neural models and, more recently, LLMs. Despite these technological advances, deployed systems often underserve their users by simply presenting a list of items, leaving the burden of exploration, comparison, and synthesis entirely on the user. This paper argues that this traditional "tool-based" paradigm fundamentally limits user experience, as the system acts as a passive filter rather than an active assistant. To address this limitation, we propose a novel deep research paradigm for recommendation, which replaces conventional item lists with comprehensive, user-centric reports. We instantiate this paradigm through RecPilot, a multi-agent framework comprising two core components: a user trajectory simulation agent that autonomously explores the item space, and a self-evolving report generation agent that synthesizes the findings into a coherent, interpretable report tailored to support user decisions. This approach reframes recommendation as a proactive, agent-driven service. Extensive experiments on public datasets demonstrate that RecPilot not only achieves strong performance in modeling user behaviors but also generates highly persuasive reports that substantially reduce user effort in item evaluation, validating the potential of this new interaction paradigm.

• The paper introduces Dr.Rec, a multi-agent framework that proactively simulates user exploration and synthesizes decision-support reports to ease cognitive load.
• The paper employs novel reinforcement learning techniques with multi-reward signals and diverse trajectory sampling to enhance behavioral modeling and recommendation precision.
• The paper demonstrates significant gains on the Tmall dataset, achieving up to 52% improvement in Recall@5 and a 49% increase in NDCG@5 over baseline methods.

Deep Research for Recommender Systems: A Multi-Agent Paradigm for Proactive, Report-Centric Recommendation

Motivation and Problem Statement

Despite significant advances in recommender system technology—including deep neural models and LLMs—fundamental limitations persist in deployed systems, which remain wedded to the traditional item-list paradigm. Current systems act as passive filters, presenting item lists for user-driven exploration, comparison, and synthesis, thereby imposing substantial cognitive and interaction costs on the user, particularly in high-involvement decision scenarios (e.g., expensive goods). The paper "Deep Research for Recommender Systems" (2603.07605) challenges this item-centered interaction model and proposes a transition to a proactive, agent-driven paradigm that autonomously simulates exploration and synthesizes decision-support reports.

Dr.Rec: Multi-Agent Deep Research Pipeline

The central contribution is the Dr.Rec architecture, a multi-agent framework that operationalizes the deep research paradigm for recommendation. It comprises two coordinated agents: a user trajectory simulation agent and a self-evolving report generation agent, as outlined below.

Figure 1: Overview of Dr.Rec. The pipeline combines autonomous user exploration simulation via RL-optimized agents with a multi-aspect, rubric- and experience-driven report generator supporting self-evolution.

User Trajectory Simulation Agent

The first agent models the entire exploration-to-decision process as a conditional generative task, using contextual and historical interactions to auto-regressively synthesize plausible future behavioral trajectories. By resorting to generative modeling with reinforcement learning (employing model-free, rule-based rewards and the GRPO objective), the agent robustly simulates the breadth and depth of potential user interaction paths, moving beyond truncated one-step prediction typical of pure discriminative models.

Core technical capabilities include:

• Session-aware trajectory tokenization: Behavioral sessions are structured to capture action-type prefixes and compress high-frequency, low-information actions.
• Multi-reward RL optimization: Training combines outcome (final purchase match), process (collaborative semantic similarity via embedding max-pooling), and constraint rewards (length and structure penalties). This multi-signal optimization addresses the limitations of outcome-only RL, providing both end-state and intermediate semantic guidance.
• Diverse trajectory sampling: Top-$p$ sampling and temperature control increase coverage of plausible user interests beyond beam search or deterministic methods, enabling the generation of high-confidence candidate exploration-to-decision sequences.

Self-Evolving Report Generation Agent

The second agent materializes the transition from an item list to a structured, interpretable report. It processes simulated trajectories to generate decision-support artifacts along multiple axes:

• Dual-channel preference modeling: User priors are factored into structured rubrics (category-attribute weights supporting long-term, stable, and controllable preference encoding) and unstructured experience (contextual, scenario-dependent cues mined from natural language and stored in vector memory).
• Multi-aspect ranking: The agent decomposes user intention into fine-grained aspects (attribute subsets), applies aspect-specific ranking using LLM-based evaluation, and consolidates these to create both overall and aspect-specific ranked lists. This supports both global and facet-level comparison and explanation.
• Structured, interpretable report synthesis: The output is a report containing the simulated trajectory, intent summary, primary and aspect-oriented recommendations, and explicit decision rationales, all grounded in traceable evidence.

Additionally, the agent is equipped with a training-free self-evolution module for both rubrics and experience memory. Online user feedback continuously informs fine-grained rubric weight updates (via best-of-$n$ attribution) and semantic experience expansion (via LLM-guided contrastive extraction), enabling lifelong, sample-efficient personalization.

Empirical Results

Rigorous evaluation—both automatic and with human annotators—demonstrates that Dr.Rec yields substantial numerical improvements in both behavioral modeling and report quality on the Tmall dataset:

• Trajectory simulation: Dr.Rec reports up to 0.1557 Recall@5 and 0.1241 NDCG@5 on purchase prediction; these correspond to up to 52% and 49% improvements over top multi-behavior baselines (MBSTR, PBAT) and reasoning-enhanced models (ReaRec, Plan-and-Solve).
• Report generation: Human and LLM evaluators converge on significant superiority in clarity, coverage, and especially novelty (77% win rate in pairwise comparison; see Figure 2), validating that Dr.Rec generates reports with more evidence-backed, non-trivial insights and reduced user cognitive effort compared to general-purpose LLMs and agent planning frameworks.

Figure 2: Win-rate analysis in pairwise report comparison reveals Dr.Rec's dominance over Plan-and-Solve, especially in Novelty and Clarity.

Ablations confirm the necessity of all core mechanisms: process and structural rewards in RL, interest decomposition, dual-channel preference representations, and self-evolution. Additional analyses identify sensitive dependencies on hyperparameters governing trajectory sampling diversity (see Figure 3), which control the trade-off between exploration coverage and semantic coherence.

Figure 3: Recall and NDCG performance on Tmall as a function of top-$p$ and temperature, highlighting the importance of controlled trajectory diversity.

Implications and Theoretical Impact

This work reframes the recommender system pipeline: rather than relying on users for iterative exploration and synthesis, intelligent agents absorb these burdens, autonomously traversing item space and providing validated, user-aligned recommendations in interpretable report form. This represents a decisive break from tool-centric paradigms and shifts towards agentic interfaces, mirroring concurrent transformations in IR (deep research) and knowledge work (LLM-based multi-agent orchestration).

Key implications include:

• Interaction cost minimization: By automating labor-intensive behaviors (e.g., attribute extraction, comparison, intent synthesis), cognitive load is transferred from end-users to the agent, with empirical reductions in interaction steps and mental workload.
• Explainability and trust: Reports codify not only the recommended items but disclose (and justify) the underlying reasoning pathways, critical for high-value or risk-averse decision contexts.
• Personalization as a continuous, real-time trajectory: The self-evolution logic enables models to rapidly adapt to shifts or drifts in user interest based on sparse feedback, without periodic retraining.
• Generalizability: The report-centric paradigm is extensible to any scenario requiring trade-off analysis or transparent rationale, including but not limited to high-priced e-commerce, B2B procurement, or professional tool selection.

Future Directions

The paradigm shift embodied in Dr.Rec opens new research avenues:

• Scalable trajectory simulation: Efficient exploration and coverage in very large item spaces remain non-trivial. Integration with learned world models, scalable RL, and amortized sampling will be important.
• Robust evaluation protocols: Traditional accuracy and ranking metrics cannot capture user experience improvements or reduction in cognitive cost, motivating new benchmarks reflecting the holistic utility of agent-generated reports.
• Dual-mode recommenders: Not all users or scenarios benefit equally from deep research; hybrid systems offering both rapid item lists and slower, in-depth reporting (with user mediation) constitute a practical deployment path.
• Factually consistent report synthesis: Reliability and calibration of LLM-generated rationales remain risk factors, and necessitate further research on hallucination mitigation, evidence tracing, and post-generation verification.

Conclusion

"Deep Research for Recommender Systems" introduces a multi-agent architecture that advances recommender system interaction from passive exposure to active, synthesized, and explainable reporting. By unifying generative trajectory simulation, aspect-level personalized ranking, and lifelong preference self-evolution, Dr.Rec establishes new SOTA performance on user behavior modeling and decision support, delivering demonstrable reductions in user effort and cognitive load. The implications are broad, positioning agent-centric, report-driven recommendation as a likely trajectory for next-generation systems that aim to be not just accurate, but deeply helpful and aligned with user decision-making.