MACDF: Multi-Agent Cognitive Decision Framework

• MACDF is a multi-layered framework featuring coordinated agents that mimic human decision-making in e-commerce searches.
• It integrates specialized agents for intent clarification, multi-source retrieval, and expert decision support to enhance recommendation accuracy.
• Empirical results show significant improvements in click-through, conversion rates, and reduced query reformulations, demonstrating its practical impact.

A Multi-Agent Cognitive Decision Framework (MACDF) is a multi-layered system architecture that redefines e-commerce search by shifting the paradigm from passive retrieval-ranking—where queries are matched directly to items—to an active, collaborative process that mirrors the cognitive decision-making stages of human shoppers. Addressing key limitations of traditional methods, MACDF integrates specialized agents for intent clarification, task decomposition, multi-source retrieval, and expert-like decision support, underpinned by collaborative memory and adaptive learning mechanisms. The architecture is engineered to synthesize raw data, interpret complex and ambiguous user queries, traverse semantic gaps, reduce decision-making costs, and deliver context-sensitive recommendations with professional guidance, thus aligning system behavior with the nuanced, multi-constraint, and reasoning-rich demands of real-world e-commerce (Zhai et al., 23 Oct 2025).

1. Architectural Structure and Agent Roles

MACDF is composed of several coordinated layers, each populated by purpose-built sub-agents that collectively execute a user’s full decision process:

• User Interaction Layer: Handles security checks, collects query input, and delivers results to the user interface.
• Multi-Agent Layer: The core, containing:
  • Leader Agent: Performs initial intent diagnosis, assesses ambiguity, launches follow-up tasks, and orchestrates sub-task scheduling using an adaptive policy:

  $$t_k = \inf(\{ t > t_{k-1} \mid \|e(t)\| > \delta \cdot e^{-\alpha t} \})$$

  where $t_k$ is the $k$-th scheduling time, $\|e(t)\|$ quantifies task error, $\delta$ is the initial threshold, and $\alpha$ reflects urgency escalation. - Guider Agent: If the intent is ambiguous, conducts slot filling and clarifying dialogs; its operation is governed by a Q-learning–inspired reward:

  $$R(s, a) = \alpha \cdot \text{Intent\_Gain} - \beta \cdot \text{Interaction\_Rounds}$$ - Planner Agent: Decomposes confirmed intent into an executable DAG $G=(V,E)$, mapping atomic tasks such as ProductSearch, WebSearch, tool invocation, and domain-specific decision modules, with both hard (symbolic) and soft (reward-penalized) constraint checking. - ProductSearch & WebSearch Agents: Combine multimodal retrieval (keyword, vector, external web), multi-source content enrichment, and LLM-driven fusion ranking employing chain-of-thought prompting. The WebSearch Agent integrates real-time data by scoring:

  $$\text{score} = \alpha \cdot \text{relevance} + \beta \cdot \text{authority} + \gamma \cdot \text{freshness}$$ - Decider Agent: Aggregates all agent outputs from the memory layer, applies multi-criteria decision analysis (e.g., Analytic Hierarchy Process, Pareto optimality), produces recommendations and full justifications, and reflects on outcomes:

  $$\text{score} = \gamma \cdot \text{Relevance} + (1-\gamma) \cdot \text{User\_Satisfaction}$$

  If the score falls below a threshold, the agent triggers re-planning.

• Memory System Layer: Centralizes context, including past subtask states, interaction history, and agent outputs, supporting robust cross-agent knowledge sharing.

• Feedback Learning Module: Continuously adapts multi-agent strategies via reinforcement learning, incorporating real user feedback (clicks, conversions) to enhance collaborative behavior.

2. Operational Paradigm Shift versus Traditional Retrieval-Ranking

Traditional e-commerce retrieval-ranking systems focus on query–item matching using textual or vector similarity. They are fundamentally limited by their inability to:

• Bridge semantic gaps in complex or multi-constraint queries, especially those involving negation or reasoning requirements.

• Reduce user effort, as users must repeatedly reformulate queries and aggregate fragmented information (e.g., by manually consolidating product specs, reviews, and external opinions), leading to high Average Reformulation Count (ARC).

• Provide professional guidance, since recommendations are derived mainly from popularity signals and fail to answer expert-level or non-standard queries.

MACDF directly mitigates these limitations by embedding layered cognitive reasoning—clarifying ambiguous needs, decomposing queries, integrating and fusing multi-source evidence, and delivering expert-informed, context-sensitive advice—resulting in a decision process akin to consulting with a domain expert rather than merely searching a repository.

3. Evaluation: Empirical Results and Performance Metrics

The framework's efficacy is demonstrated through extensive offline and online evaluation:

• Offline Benchmarks (ECCD-Bench):

  • For negation intent queries, MACDF achieved an Accuracy@Top5 (ACC@5) of 0.86 compared to 0.07–0.12 in baseline systems.
  • User Demand Satisfaction (UDS) was significantly improved, confirming that the recommendations better matched user expectations for reasoning-rich and consultative queries.
• Online A/B Testing on JD Platform:
  • User Click-Through Rate (UCTR) increased by +3.9%.
  • User Conversion Rate (UCVR) rose by +6.5%.
  • Average Reformulation Count (ARC) decreased by 7.7%, evidencing lower user friction and higher first-pass relevance.

MACDF’s layered reasoning leads not only to higher recommendation accuracy but also to improved user interaction quality and measurable business impact.

4. Cognitive Mechanisms and Agent Collaboration

A central feature of MACDF is the explicit modeling of cognitive processes analogous to human decision-making:

• Hierarchical Reasoning: Ambiguous queries are handled through hierarchical slot filling and clarification before task decomposition, mirroring human inquiry.
• Directed Acyclic Task Scheduling: Agents execute and synchronize sub-tasks via the DAG, supporting flexible parallel and sequential planning.
• Multi-Agent Consensus: The Decider Agent employs multi-criteria analysis to reconcile conflicting evidence, leveraging both analytic hierarchy and Pareto optimality principles.
• Reflection and Adaptation: Built-in reflection mechanisms (e.g., composite scoring of relevance and satisfaction, with auto-triggered re-planning) ensure continual self-assessment and refinement of recommendations.

These mechanisms collectively facilitate emergent intelligence, enabling the multi-agent system to adapt in real time to user-specific context and query complexity.

5. Applications and Implications

MACDF is primarily applied to next-generation e-commerce recommendation, where it:

• Guides decision-making in high-stakes or high-ambiguity product queries, integrating structured information (product specs), unstructured content (reviews, forums), and real-time web data.
• Delivers expert-level advice previously unattainable with standard retrieval—such as handling negations (e.g., “not made in country X”), elaborate constraints, or novel product categories.
• Reduces cognitive burden and improves satisfaction, as validated by increased conversion and click metrics and decreased reformulation counts.
• Serves as a blueprint for cross-domain decision support systems (e.g., healthcare, finance), where multi-agent collaboration, continuous knowledge integration, and cognitive alignment with human reasoning are critical.

6. Limitations and Future Directions

While providing marked improvements over classical systems, MACDF introduces notable computational burdens:

• Latency: The total response time (≈15 seconds) and time-to-first-token (1.8 seconds) are significantly higher than traditional approaches (~0.8 seconds). Optimization strategies such as parallelization, pre-computation, and model compression are proposed to address these concerns.
• Need for Extended Evaluation: Current assessments focus on short-term engagement and conversion. Future research will involve long-term impact studies (user retention, lifetime value) and more granular ablation studies to attribute performance gains to specific agent modules and collaborative strategies.
• Ablation and Interpretability: Systematic studies are needed to further elucidate which agent roles and interaction policies contribute most to overall effectiveness, informing streamlined and context-adapted MACDF variants.

7. Significance in Multi-Agent Cognitive Systems

MACDF operationalizes recent advances in multi-agent systems and cognitive modeling, demonstrating that expert-inspired, multi-stage agent orchestration can fundamentally redefine user experience and outcome quality in search-based applications. By bridging the semantic and operational gaps of retrieval-ranking, MACDF’s architecture establishes a path forward for multi-agent cognitive decision frameworks that are robust, adaptive, and capable of handling real-world complexity in e-commerce and beyond (Zhai et al., 23 Oct 2025).